검색
검색 팝업 닫기

Ex) Article Title, Author, Keywords

Article

Split Viewer

Original Article

Progress in Medical Physics 2024; 35(4): 178-204

Published online December 31, 2024

https://doi.org/10.14316/pmp.2024.35.4.178

Copyright © Korean Society of Medical Physics.

Development of an Instantaneously Interpretable Real-Time Dosimeter System for Quality Assurance of a Medical Linear Accelerator

Dongyeon Lee1,2 , Sung Jin Kim2 , Wonjoong Cheon3 , Hyosung Cho1 , Youngyih Han2,4

1Department of Radiation Convergence Engineering, Yonsei University, Wonju, 2Department of Radiation Oncology, Samsung Medical Center, Seoul, 3Department of Radiation Oncology, Seoul St. Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, 4Department of Health Sciences and Technology, SAIHST, Sungkyunkwan University, Seoul, Korea

Correspondence to:Youngyih Han
(youngyih@skku.edu)
Tel: 82-2-3410-2604
Fax: 82-2-3410-2619

Received: September 23, 2024; Revised: December 13, 2024; Accepted: December 13, 2024

This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Purpose: Modern radiotherapy delivers radiation doses to targets within a few minutes using intricate multiple-beam segments determined with multi-leaf collimators (MLC). Therefore, we propose a scintillator-based dosimetry system capable of assessing the dosimetric and mechanical performance of intensity-modulated radiotherapy (IMRT) and volumetric-modulated arc therapy (VMAT) in real time.
Methods: The dosimeter was equipped with a scintillator plate and two digital cameras. The dose distribution was generated by applying deep learning-based signal processing to correct the intrinsic characteristics of the camera sensor and a tomographic image reconstruction technique to rectify the geometric distortion of the recorded video. Dosimetric evaluations were performed using a gamma analysis against a two-dimensional array and radiochromic film measurements for 20 clinical cases. The average difference in the MLC position measurements and machine log files was tested for the applicability of the mechanical quality assurance (QA) of MLCs.
Results: The agreement of the dose distribution in the IMRT and VMAT plans was clinically acceptable between the proposed system and conventional dosimeters. The average differences in the MLC positions for the IMRT/VMAT plans were 1.7010/2.8107 mm and 1.4722/2.7713 mm in banks A and B, respectively.
Conclusions: In this study, we developed an instantaneously interpretable real-time dosimeter for QA in a medical linear accelerator using a scintillator plate and digital cameras. The feasibility of the proposed system was investigated using dosimetric and mechanical evaluations in the IMRT and VMAT plans. The developed system has clinically acceptable accuracy in both the dosimetric and mechanical QAs of the IMRT and VMAT plans.

KeywordsDosimeter, Real-time, Deep learning, Dose rate, Multi-leaf collimators

Modern radiotherapy techniques allow normal tissues to receive a minimal dose of radiation while maintaining a prescription dose conformal to the target volume. Specifically, intensity-modulated radiotherapy (IMRT) [1] and volumetric-modulated arc therapy (VMAT) [2] deliver radiation doses to a target within a few minutes using intricate multiple beam segments determined with jaw and multi-leaf collimators (MLCs). The combination of multiple beam segments allows the radiation dose to be more focused on the target as compared with the three-dimensional conformal radiotherapy (3DCRT) [3-5]; however, the complexity of dose modulation in IMRT and VMAT requires a more suffocating validation process as compared with 3DCRT [6-8]. Therefore, the task group (TG) of the American Association of Physicists in Medicine (AAPM) has published reports on the patient-specific quality assurance (QA) of IMRT and VMAT plans; these reports have been widely adopted by radiation therapy institutes to reduce the uncertainty of IMRT and VMAT [9-11].

Patient-specific QA can be performed by comparing the measured dose distributions with the dose map computed from the treatment planning system (TPS). According to a recent survey, 80.6% of the sites responded that they conducted patient-specific QA with planar measurements to verify the delivered dose [12]. Planar dosimetry can be performed using a radiochromic film, an electronic portal imaging device (EPID), a two-dimensional (2D) diode array, and a 2D ionization chamber array. Radiochromic films have advantages, including a high spatial resolution and convenience of employment [13]. However, they cannot be reused and are unsuitable for real-time dosimetry because of the saturation time required to stabilize the discoloration [14]. The EPID has shown potential to replace film dosimetry [15,16]. Generally, dosimetry using the EPID was performed by applying the water dose conversion to the acquired radiograph image [17-21], and the properties of the EPID allowed performance of the machine-independent measurements. However, the instability in dosimetry has been presented, since the EPID was originally designed for imaging applications. Furthermore, structural limitations, including increased non-uniform backscatter due to the supporting arm, caused a non-linear response to radiation [22-24].

The continuously varying dose rates of VMAT allows the MLC to be modulated at a high speed such that the delivery time of VMAT is shorter than that of IMRT. Additionally, a continuously varying dose rate contributes to improved beam flatness and stability as compared to a fixed dose rate [25]. However, the high speed of the MLC increases the positioning errors of leaves [26,27]; therefore, the instant dose rate and accuracy of the MLC position are important components of VMAT quality.

A scintillator plate-based dosimeter with a camera was first developed in 1980 by Baily et al. [28] and improved by Boon et al. [29] for the direct measurement of the radiation dose in a water phantom. The characteristics of the scintillator with respect to linearity, repeatability, and stability were evaluated, and its feasibility as a dosimeter was proved [30]. Camera image sensors, including complementary metal-oxide-semiconductors (CMOSs) and charge-coupled devices (CCDs), are damaged when the camera is in the radiation field, which can increase the dark current and hot pixels in the captured image [31-33]. Thus, in the conventional scintillator dosimeter, the camera was mounted on the side and a mirror was installed to protect the image sensor from the primary beam and prevent geometric distortion of the recorded image by reflecting the scintillation [34-37]. However, the built-in mirror increased the dosimeter volume, and the commercial scintillator dosimeter, for example, Lynx PT (IBA Dosimetry) has dimensions of 360×370×600 mm3. Although the accessibility of scintillator dosimeters is limited due to their excessive volume, recent advances in signal processing, which enable cameras to capture images with a high temporal resolution, have shown promise as verification devices for VMAT [38].

In the present study, we developed a mirror-less real-time scintillator dosimeter for the QA of a medical linear accelerator (LINAC), specifically for IMRT and VMAT plans, which can detect the varying dose rates of delivering beams. In this article, the IMRT was defined as a plan with a static MLC during beam delivery, whereas the VMAT was defined as a plan with a dynamic MLC. The dosimeter was composed of a scintillator plate and two CMOS-based cameras without a mirror. The proposed system was capable of measuring the dose distributions at 60 frames per second (FPS), which corresponds to a sampling rate of 16.67 ms. The high temporal resolution provided a function for interpreting the instantaneously modulated dose distribution in the media. Specifically, the proposed system performed a real-time analysis of VMAT, which could predict the dose rates of the LINAC and track the MLC positions in multiple beam segments through highly sampled dose maps. Removing the mirror improved the usability by reducing the weight to 4.08 kg. The characteristics of the developed system were analyzed and the dosimetric and mechanical performance were verified.

1. Design of the dosimeter

The proposed dosimeter was composed of a scintillator plate and two digital cameras with CMOS image sensors in an aluminum-frame-based dark box (Fig. 1). A scintillator plate was placed on the top of the dark box, and two digital cameras were installed at the bottom, tilted to face the scintillator plate. Both cameras had the same tilting angle of 57°. The dosimeter had a volume of 316×606×306 mm3 and a weight of 4.08 kg. All sides of the dosimeter were covered with a black paper to block the external light.

Figure 1.Experimental setup of the developed dosimeter with a scintillator plate and complementary metal-oxide-semiconductor based digital cameras.

The PI-200 (Mitsubishi Chemical) set on top of the dark box was a flat-plate-type inorganic scintillator with a size of 300×300×0.9 mm3. It includes three layers of protective phosphor and supporting materials. The phosphor layer was composed of terbium-doped gadolinium oxysulfide, which is widely used in medical imaging systems.

A DSC-RX100M7 (Sony Group Corp.) digital camera recorded the emitted light from the scintillator plate (volume of 42.8×101.6×58.1 mm3 and a weight of 302 g). The recorded video had high-resolution images with pixel dimensions of 1,920×1,080 and a frame rate of 60 FPS, equivalent to 16.67 ms in a single frame. The cameras were operated using remote controllers during irradiation. After recording, the video from the two digital cameras was transmitted to a computer for signal processing, as described below.

2. Signal processing of the proposed system

The light emitted from the scintillator was recorded by a camera using a CMOS image sensor. The main peak in the spectrum of the emitted scintillation was approximately at 530 nm, and the recorded videos are mainly presented in green. Therefore, we performed signal processing using the green channel of the recorded video to convert the recorded video into a dose map. Fig. 2 shows the signal processing of the proposed dosimeter system.

Figure 2.Simplified flow chart of the signal processing in the developed system. Output correction is made in the signal processing. IMRT, intensity-modulated radiotherapy; VMAT, volumetric-modulated arc therapy; FT, Fourier transformed.

First, when the recorded video was input, the video was summed along the frames, and a threshold value was applied to the summed video to obtain a beam signal. Thereafter, the masks of the radiation fields were extracted from the video of the IMRT and VMAT plans through segment and control point (CP)-based processing, respectively. Given that IMRT and VMAT have different mechanisms of dose delivery, we divided the signal processing in two ways. Specifically, the recorded video of the IMRT was disturbed by aliasing between the frame rate of the camera and the fixed beam pulse rate of the LINAC in each segment. Therefore, we applied an iterative Fourier transform (FT)-based aliasing correction to the recorded IMRT video before mask extraction. A detailed description is provided in Section “Iterative Fourier transform-based aliasing correction”. Thereafter, the IMRT mask was generated after applying a threshold of 0.45 to the normalized segment image. The threshold value for the IMRT video mask extraction was selected empirically. The segmented image was acquired by accumulating the video frames recorded at each subfield of the IMRT plan. Fig. 3a shows an example of the recorded video images and the extracted masks from a single segment of the IMRT plan. However, VMAT modulates the radiation doses with consecutively varying radiation fields built from the jaw and MLC, and it is controlled in the LINAC through CPs. The recorded video of the VMAT plan had a different radiation field in each frame corresponding to a CP. However, a single frame of the VMAT plan recording includes grain noises because of the high international organization for standardization (ISO) value setting of the digital cameras [39,40], whereas the recorded video from the IMRT plan was free from grain noises due to the accumulated frames. Consequently, the mask generated by applying a threshold to a single frame of the VMAT plan was incompletely binarized, as shown in Fig. 3b. Therefore, we applied deep learning-based mask extraction to the VMAT plan’s recorded video to overcome this limitation. The details of the deep learning-based mask extraction of the VMAT are described in Section “Video mask extraction of the VMAT plan”.

Figure 3.Extracted mask from the recorded videos of the (a) IMRT and (b) VMAT plans by the thresholding method. The mask of the IMRT plan was extracted from the segmented image accumulated during the beam-on time, whereas that of the VMAT plan was obtained from the single-frame image corresponding to the CP. IMRT, intensity-modulated radiotherapy; VMAT, volumetric-modulated arc therapy; CP, control point.

After the mask extraction, a signal correction was applied to the video image. Camera image sensors, including CMOS and CCD, have intrinsic sensitivity to brightness according to the ISO value. The sensitivity of the camera sensor to a certain ISO value is indicated by an S-shaped curve for the dynamic range of light [41,42]. Generally, camera sensors with higher ISO values respond to photons more quickly as compared to those with lower ISO values. Therefore, cameras in dark environments often use high ISO value settings. In the present study, an ISO value of 12,800 was adopted to record the light of the scintillator plate in a dark box. However, because of the intrinsic response of the camera sensor to brightness, the output factors for several radiation field sizes measured with the proposed dosimeter were different from the reference values measured by the ionization chamber (CC13; IBA Dosimetry), as shown in Fig. 4. The output factor of the proposed dosimeter before the signal correction was observed to increase sharply as compared to that of the reference in large-sized radiation fields.

Figure 4.(a) Comparison of the output factors between the ionization chamber and proposed system before the output correction and (b) correction curve of the proposed dosimeter in the solid water phantom thickness of 5 cm.

A cubic function-based correction curve was employed in the recorded video to mitigate the discrepancy between the camera and ionization chamber measurements. The output factor is dependent on the size of the radiation field; thus, we computed the equivalent field size (Acomputed) from the extracted mask (mask) as follows:

Amask= iX jYmaski,j,
Acomputed=Amask1 +Amask2 2Cmask,

where i and j are the pixel positions of the extracted mask in the vertical and horizontal directions, respectively. Furthermore, A(mask) is the number of pixels selected in the extracted mask from the recorded video, and it was computed from the binarized mask image. The Cmask is the positive constant that converts the pixel size to the equivalent field size. A digital image is presented by numerous pixels of a specific size. The field size of the region of interest can be calculated using the information of the pixel size and number of pixels. Generally, the selected pixel is represented by one and the background is represented by zero in the binary image, such as mask. Therefore, the number of pixels selected in the mask can be computed by adding up the total pixels of the image [Eq. (1)]. However, the recorded videos had geometric distortion due to the position of the digital cameras. In the present study, we used the average number of pixels in two masks (mask1 and mask2) extracted from the recorded videos of two different digital cameras to mitigate the effects of geometric distortion in the images [Eq. (2)]. The equivalent field size was computed after applying the square root to the mean and dividing by the pixel size correction factor Cmask. The linear response between the computed and actual field sizes was observed (Fig. 5).

Figure 5.The linear response between the computed and actual field sizes. The computed square root field size was calculated from the extracted masks of the recorded videos.

Thereafter, the correction curve was estimated using a cubic function based on the curve fitting function of MATLAB (R2021a, MathWorks), and the computed radiation field size from the extracted mask and output factor ratio of the reference to the measurement were used as the input parameters. The output factor correction was performed using the equivalent field size for each frame of the recorded video. The setup had a source-axis distance of 100 cm to a 5-cm deep solid water phantom, where the signal correction curve was estimated. The distributions under different setup conditions from the aforementioned state can be measured after applying signal correction, which is calculated under the setup condition to be observed.

The proposed dosimeter required space under the scintillator plate to capture the light produced as a result of the interaction between the scintillator material and radiation. This structural design makes it difficult to ensure the contribution of the backscatter generated from the materials behind the scintillator plate. Generally, a backscatter thickness of ≥6 cm is required for a 6-MV photon beam [43,44]. However, the signal correction of the proposed system was used to rectify every error caused by the camera sensor and structural limitations, including the backscattering deficiency. After the signal correction, a geometric correction was performed to recover the distortion caused by the position of the digital cameras. The details of the geometric correction are described in Section “Iterative reconstruction technique-based geometric correction”.

The prescribed signal processing procedure for the IMRT and VMAT plans are summarized in Fig. 2.

3. Video mask extraction of the volumetric modulated arc therapy plan

The proposed dosimeter system corrected the output factor errors using the correction curve based on the estimated size of the radiation field from the mask. Thus, it is important to acquire the exact mask of the radiation field from the recorded video images. However, the VMAT plan’s recorded video had a limitation in mask generation using the thresholding method, as described in Section “Video mask extraction of the VMAT plan”. Recently, image processing based on deep learning has demonstrated outstanding performance in the medical field, including image segmentation [45,46]. In particular, the usability of the UNet-based mask extractor has been proven [47]. However, the enlarged receptive field of UNet with pooling layers can lose the details of an image presented at a high frequency [48,49]. Therefore, we applied a cycleGAN-based mask extractor [50-52] to the VMAT plan’s recorded videos.

A simplified diagram of the cycleGAN-based VMAT mask extraction algorithm is shown in Fig. 6. In the proposed algorithm, the cycleGAN model included the following two trainable components: generators and discriminators. The generators and discriminators of cycleGAN work complementarily to achieve convergence during the training process. The generator was trained to produce fake images by reflecting the relationship between the masks (IMask) and video images (IVideo). Specifically, IMask and IVideo were paired inputs and labels, respectively. The discriminator was trained to evaluate whether the images produced by the generator were similar to the label images. In the present study, the generator was composed of nine residual blocks based on the ResNet architecture [53], and the discriminator was adopted from the PatchGAN algorithm [54].

Figure 6.A simplified diagram of mask extraction in the VMAT plan’s recorded video images employing the cycleGAN model. IMRT, intensity-modulated radiotherapy; TPS, treatment planning system; DICOM-RT, digital imaging and communication in medicine of radiotherapy; VMAT, volumetric-modulated arc therapy.

The proposed network was trained with the video images of the IMRT plans and masks were calculated from the digital imaging and communication in medicine of radiotherapy (DICOM-RT) files generated by the TPS (Eclipse ver. 15.5; Varian Medical Systems) at Samsung Medical Center. Generally, the DICOM-RT file includes geometric and dosimetric data specifying the course of radiotherapy, such as the beam angles, collimator openings, and beam modifiers. Therefore, the jaw and MLC positions can be specified from the DICOM-RT file. After extracting the positions of the jaw and MLC, the MLC mask can be generated by calculation using the geometric properties of the medical LINAC. A video mask for training the cycleGAN-based mask extraction algorithm was produced by applying the forward projecting operator (ALumi), described in Section “Iterative reconstruction technique-based geometric correction”, to the MLC mask. The accumulation process for acquiring the segmented images was ignored to generate CP-like images from the IMRT plan’s recorded video.

4. Iterative Fourier transform-based aliasing correction

We used an X-ray beam with an energy of 6 MV from a me­dical LINAC (TrueBeam STx; Varian Medical Systems) at various dose rates. The dose rate was defined as the monitor unit (MU) per minute of the medical LINAC. In general, the dose per pulse of a medical LINAC is constant, and the dose rate is varied by modifying the beam pulse rate [55,56]. The TrueBeam STx used in this study has a beam pulse rate of 60 Hz at a dose rate of 100 MU/min. The beam pulse frequency was sequentially increased to 120 Hz at 200 MU/min, 180 Hz at 300 MU/min, 240 Hz at 400 MU/min, 300 Hz at 500 MU/min, and 360 Hz at 600 MU/min. Therefore, we set the camera frame rate to 60 FPS to cover all configurable dose rates in the TrueBeam STx. The scintillation signal input to the camera can be periodically attenuated owing to the sampling frequency of the image sensor, as shown in Fig. 7a, which is referred to as aliasing [57]. As a result, the signal from a single frame had a linear relationship with the dose rate; however, the dose rate response of the accumulated frames was indicated differently even when the same dose was irradiated. (Fig. 7b) Thus, we applied an iterative FT-based aliasing correction method to the IMRT signals, as shown in Fig. 8. First, the signal from the IMRT was FT to separate the input signal into real and imaginary parts in the frequency domain. The zero-frequency component (FT(x0)) was extracted for use as a criterion for the aliasing correction. Thereafter, the signal (xk) that contributed to building the aliasing in the input signal was selected from the real and imaginary parts as follows:

Figure 7.Comparison of the following signals measured by the proposed dosimeter among the different dose rate settings of the TrueBeam STx (Varian Medical Systems): (a) periodically attenuated signals due to the interference of the sampling frequency between the linear accelerator and cameras, (b) response to the dose rates in the single frame and accumulated frames, and (c) calibrated signal. MU, monitor unit.

Figure 8.Flowchart of the iterative Fourier transform (FT)-based aliasing correction algorithm. DC, direct current.

xk=10k<RealFTxk<10k+1,

where the initial k was set to 1 in. the iteration. The signals selected from the real and imaginary parts were attenuated by 10k to relive the aliasing of the original signal. The real and imaginary parts were successively updated (x*) until k satisfied the following criteria:

k<log10RealFTx0

Finally, the corrected signal was obtained by applying an inverse FT to x*, as shown in Fig. 7c.

Generally, the VMAT delivers the radiation with a continuously variable dose rate during the beam-on time, and the continuously variable dose rate can be indicated as a multi-band signal. The measured signals at different dose rates of the IMRT, including aliasing (Fig. 7a), had low frequencies, and these were negligible at continuously variable beam pulse rates. Therefore, we applied the FT-based aliasing correction step only to the IMRT plans.

5. Iterative reconstruction technique-based geometric correction

The side-positioned cameras caused geometric distortions, including prospective distortions, owing to the tilting angles, as shown in Fig. 9a. Thus, we applied an iterative reconstruction technique (IRT) to the recorded video, allowing geometric distortion to be corrected, even in wide radiation fields. First, we assumed that the proposed dosimeter was a digital tomosynthesis model, as shown in Fig. 9b. Generally, computed tomography (CT) systems can be restored using IRT, a mathematical process that generates tomographic images from the projection data acquired at many different angles with repetitive calculation. In this study, we applied a fast iterative shrinkage-thresholding algorithm (FISTA) [58] to correct the geometric distortion of the proposed dosimeter. The details of the parameters used in the IRT-based geometric correction are listed in Table 1.

Table 1 Parameters used in the IRT-based geometric correction for dosimetric and mechanical QA

ParameterDosimetric QAMechanical QA
Source-to-detector distance (mm)543.72543.72
Source-to-object distance (mm)357.71357.71
No. of projections22
Pixel size (mm)0.30.3
Pixel dimension1,920×1,0801,920×1,080
Voxel size (mm)10.1
Voxel dimension300×3003,000×3,000
Reconstruction algorithmFISTAFISTA

Figure 9.(a) Examples of the recorded video with geometric distortion (left) and corresponding dose map corrected with the tomographic image reconstruction technique (right). The geometric distortion of the recorded video was corrected by considering the proposed dosimeter as a (b) digital synthesis model.

The mathematical settings for the description of the FISTA framework in a general CT system were prepared as follows. The original image vector to be reconstructed (x) and the measured projection vector (b) are represented by

x=x1,x2, ...,xNT
b=b1,b2, ...,bMT

where N is the number of voxels, M is the total number of sampling points in the projection data, and the superscript T is the transpose operator. The system matrix is defined as the matrix including the weighting of the interaction model between every image voxel point and all the different rays in the projection:

A=aij,  i=1, 2, ...,M and j=1, 2, ...,N

Using these definitions, the measurement equation is described as follows:

Ax=b

The linear equation Eq. (4) has an infinite solution under the condition M<N, which is physically identical to the CT reconstruction. Thus, the regularization method with l1 norm, which is less sensitive to the sharp edges of the images as compared to l2 regularization, is employed to relieve the instability of Eq. (4) and for optimization. The l1 regularization method enables Eq. (8) to be solved by reducing N to M by exploiting the sparsity of the optimal solution (x*):

x*=minxFxfx+gx
fx=12Axb22
gx=λx1

where f is a smooth convex function and g is a non-smooth convex function owing to the l1 term. The quadratic approximation (QL) is derived from F at a given point y to optimize the objective function by reducing the impact of the non-smooth convex function g. Thereafter, the process of searching the x* as a convex optimization problem (pL) can be rewritten:

QLx,y=fy+xy,fy+L2xy2+gx
pLx,y=argminxQgx+L2xy1Lfy2

where L is a Lipschitz constant and Q is the set of feasible x.

The convex optimization problem described in Eq. (9) can be solved approximately but efficiently using the accelerated proximal gradient descent method [59,60]. In the CT reconstruction process, the system matrix is repeatedly expanded to calculate the forward and backward projections in each iteration loop.

The conventional system matrix (A) for CT reconstruction was computed using Siddon’s algorithm [61]. Siddon’s algorithm calculates the projection data using the intersection length along the voxels of the image volume. However, the proposed dosimeter system measures the emitted light of the scintillator plate, which is recorded using a digital camera following the luminescence intensity function (B):

B=I4πr2

where I is the intensity of the light from the source and r is the distance between the source and measurement point. Therefore, we developed a luminescence intensity-based system matrix (ALumi) to reconstruct the radiation field image and replace A with ALumi in the FISTA framework.

ALumi=14πI1a1,12I2a1,22INa1,N2I1a2,12I2a2,22INa2,N2I1aM,12I2aM,22INaM,N2,

where aM,N is the distance between the center of the image voxel and camera image sensor.

6. Multileaf collimator mask generation

The dose maps measured by the developed dosimeter were built from the jaw and MLC of the LINAC. Traditionally, the radiation field size is defined as the 50% isodose line of the beam profiles at the maximum dose. The MLC positions in the measured dose maps can be predicted using the binarized mask after applying a simple threshold, as shown in Fig. 10.

Figure 10.An example of the multi-leaf collimators (MLC) mask generation in the intensity-modulated radiotherapy (IMRT) plan.

Thresholding-based mask extraction for the prediction of the MLC position in VMAT plans was unable to produce binary images similar to the real MLC position maps because of the increased noise in a single frame, as shown in Fig. 11. Therefore, we applied wavelet-assisted UNet (WaveUNet) to generate MLC masks from the reconstructed CP dose images (Fig. 12).

Figure 11.An example of the multi-leaf collimators (MLC) mask generation in the volumetric-modulated arc therapy (VMAT) plan. CP, control point; WaveUNet, wavelet-assisted UNet.

Figure 12.A simplified diagram representing the network structure of the WaveUNet for extracting the multileaf collimator mask from the control point dose image in the volumetric modulated arc therapy case.

A single scale of the 2D discrete wavelet transform without a diagonal detail [62] was applied to the reconstructed CP dose images before being feedforward to the network. The input wavelet-transformed images were downsized through the pooling layers in the encoding part of the WaveUNet and upsized through the unpooling layers in the decoding part. The WaveUNet includes a shortcut connection indicated by dotted arrows in Fig. 12 to include features of the encoding part in the decoding part, as a conventional UNet model [47]. The WaveUNet model was then optimized using the binary cross-entropy loss function (LBCE) as follows:

LBCE=IlabellogIpred+1Ilabellog1Ipred

where Ilabel represents the paired-label MLC masks computed from the DICOM-RT files and Ipred represents the synthesized binary images when the CP-like dose images from the IMRT plan are entered into WaveUNet. A total of 8,948 paired images were prepared to train the WaveUNet model, and the Adam optimizer [63] optimized the network at a learning rate of 1×10−4 during the training process of 1,000 epochs.

7. Prediction of the linear accelerator dose rate

Given that the developed system measured the absolute dose distributions at a high sampling rate, the dose rate of the LINAC was predictable. The proposed dosimeter measured the dose distribution every 16.67 ms during the beam-on time. First, instantaneous doses were extracted from all the frames recorded during the beam-on time. The LINAC dose rate was then estimated after applying compensation for the measured dose rates. X-rays were delivered from the LINAC during the beam-on time at a constant dose rate; however, the measured dose rates were different because the radiation field sizes and doses were modulated by the jaw and MLCs. Thus, the dose rate of the LINAC was estimated after applying compensation to the instantaneous doses. In the signal correction process, we acquired the reference output factors using the ionization chamber and the equivalent radiation field sizes computed from the extracted video masks. The compensation process was performed by dividing the instantaneous doses by the calculated output factors from the function derived from the computed field sizes and output factors. The predicted LINAC dose rates from the developed dosimeter were compared with those calculated from the TrueBeam trajectory log files.

8. Analysis of the dose map reconstruction quality

In a previous study, an image registration technique was applied to correct the geometric distortion of a scintillator dosimeter using a single camera [64]. The registration-based correction method restored the geometric distortion of the recorded video image by applying the affine transformation, including shear. The registration method was applied to the dual camera-based proposed system to compare the accuracy of the geometric distortion correction with that of the IRT-based method. The IRT-based method used to correct the geometric distortion derived from tilted cameras is widely used in CT reconstruction. Video images from the two cameras were used to reconstruct the dose distribution image of the developed dosimeter, which corresponds to two projection data points in the tomographic system. The other reconstruction algorithms for the analytical and iterative methods were implemented for comparison. Filtered back projection (FBP) and maximum-likelihood estimation–maximization (MLEM) [65] were used as the analytic and iterative methods, respectively, for the comparison of the reconstruction algorithms. The correction of the geometric distortion in the developed dosimeter employing the tomographic image reconstruction technique was performed using a MATLAB code accelerated by a graphic processing unit for fast calculation [66].

9. Validation of the dosimetric performance of the proposed system

The proposed dosimeter system was evaluated in terms of its general, dosimetric, and mechanical performances. The output factor, linearity, and dose-rate dependency were assessed using general performance tests. In the output factor test, we exposed 105 MU equivalent to 100 cGy with field sizes of 3×3, 5×5, 10×10, 15×15, 20×20, 25×25, and 30×30 cm2. The linearity of the developed dosimeter was evaluated using a field size of 10×10 cm2 and delivered doses of 25, 50, 100, 150, 200, and 250 cGy. A fixed dose rate of 400 MU/min was used for the output factor and linearity tests. The reference values for the output factor and linearity test were measured using an ionization chamber (CC13). The dose-rate dependency of the proposed system was observed by irradiating X-ray beams with a dose of 300 cGy and dose rates of 100, 200, 300, 400, 500, and 600 MU/min. In the dose-rate dependency test, the EBT3 film (Ashland) was used as a reference to minimize any influence associated with the sampling frequency issue caused by the digitized dosimeters.

The commercial gamma analysis software (RIT113 ver. 6.2; Radiological Imaging Technology Inc.) was used to evaluate the dosimetric performance of the proposed dosimeter [67,68]. The gamma passing rate was computed to quantify the coincidence between the dose distribution measured by the dosimeter and the dose map acquired from conventional dosimeters, such as a 2D array dosimeter and radiochromic film. The 2D array dosimeter was chosen as it was the sole option capable of instantaneously capturing the modulated dose distributions from the VMAT plans. The IMRT and VMAT plans were generated from 10 clinical cases with five different treatment sites (i.e., brain, head and neck, lung, abdomen, and pelvis); therefore, a total of 20 plans were used in the present study. Given that the proposed system measured the dose distribution by placing the dosimeter on the couch, the clinical plans were recalculated through TPS under the condition of delivering doses from a fixed gantry and collimator set at zero degrees. Specifically, recalculation of the clinical plans to a phantom was performed after anonymizing the individual patient data.

The 2D array dosimeter (MatriXX ResolutionTM; IBA Dosimetry) measured the dose distribution with a pixel spacing of 6.5 mm. The active measurement area of the MatriXX ResolutionTM was 247×247 mm2, and the measured dose map had a resolution of 39×39 pixel. Therefore, we applied linear interpolation to expand the resolution of the MatriXX ResolutionTM dose map to 247×247 with a pixel size of 1×1 mm2. The dose distribution measurement using MatriXX ResolutionTM was performed at a sampling rate of 20 ms to verify the real-time dosimetric performance of the proposed system. For the EBT3 film analysis, digitization was performed with the 75-dots-per-inch scanning option using a 11000XL EPSON flatbed scanner (EPSON). Given that the dose map measured by the proposed dosimeter had a different image size as compared to that of MatriXX ResolutionTM, the dose map of the proposed system was cropped to 247×247. Furthermore, the scanned EBT3 film was rescaled to have a uniform resolution as the dose map of the MatriXX ResolutionTM. Thereafter, co-registration, including image rotation, was applied between the dose map of the proposed system and the measurements of MatriXX ResolutionTM and EBT3 film. Finally, a gamma analysis was performed using the concept of the global gamma through the commercial software RIT113. The real-time gamma analysis for a single CP of the VMAT plan was performed with the extracted CP dose maps, considering the different sampling rates between the developed system and MatriXX ResolutionTM. Specifically, we selected a single CP dose map with a time difference of <0.1 ms between the developed dosimeter and MatriXX ResolutionTM.

10. Validation of the mechanical performance of the proposed system

The quantitative evaluation of the mechanical performance of the developed dosimeter was conducted using the picket fence test [69] recommended by TG-142 to individually validate the alignments of the MLC leaves [10]. The nine narrow bands had a gap width of 10 mm and an interval of 15 mm. TrueBeam STx modulated the radiation fields with 60 pairs of MLC leaves affixed to two opposing parallel carriages (banks A and B). The picket fence test was performed four times with two repetitions each from bank B to bank A and bank A to bank B directions; the test from the bank A to bank B direction was defined as an inverse trial. The MLC positions of the clinical cases were also predicted using the thresholding-based MLC mask in IMRT and the WaveUNet-based MLC mask in VMAT.

The accuracy of MLC positioning was evaluated by comparing the predicted MLC positions in the proposed system and recording the MLC positions in the log files. The log file, which is a record of the LINAC data including the gantry angles, MLC positions, and delivered MU, is stored after treatment completion, and it has been reported to provide accurate MLC positions through various analyses [70-72]. Specifically, the developed dosimeter recorded the dose distribution every 16.67 ms, and the TrueBeam trajectory log file was written with a sampling rate of 20 ms. Therefore, we synchronized the time stamps of the developed dosimeter and log file using linear interpolation to compare the MLC positions in the same phase:

post=posLogt20×20+posLogt20×20posLogt20×20×tt20×2020,

where pos(t) is the MLC position computed from the log file (posLog) at the time of the single-frame dose map in the proposed dosimeter (t). The mechanical performance was validated using the mean absolute error (MAE), which measures the average magnitude of the MLC positioning error between the proposed system and the log file used as a reference:

MAE=i=1npost,ipospredt,in

where i is the leaf number and pospred is the predicted MLC position from the developed dosimeter.

1. Quality of the reconstructed dose map

The examples of the corrected dose distributions obtained by using the registration and IRT-based methods are shown in Figs. 13 and 14. The plots from Figs. 13 and 14 demonstrate the measured beam profiles along AB¯¯ and CD¯¯ as shown in Figs. 13a and 14a, respectively. The registration technique restored the geometric distortion of the proposed dosimeter, similar to the EBT3 film, in a small-sized radiation field. However, the corrected radiation fields were expanded when the size of the radiation fields was increased (Fig 13d), because the registration-based method forcibly stretched the skewed image using the transformation matrix computed during the calibration process [64]. Specifically, the transformation matrix was derived from the four vertices selected by the observer in the checkerboard image within the field of view of the scintillator detector. The IRT-based correction method preserves the reconstruction accuracy of the dose distributions, even for large radiation fields.

Figure 13.A comparison of the (a) correcting methods for geometric distortion and dose profiles in field sizes of (b) 5×5, (c) 10×10, and (d) 20×20 cm2 measured along AB¯¯.

Figure 14.A comparison of the (a) reconstruction algorithms for geometric distortion and dose profiles in field sizes of (b) 5×5, (c) 10×10, and (d) 20×20 cm2 measured along CD¯¯. FBP, filtered back projection; MLEM, maximum-likelihood estimation–maximization.

The analytic reconstruction algorithm, FBP, severely degraded images with streak artifacts induced by theoretically insufficient angular sampling. The iterative reconstruction algorithm, MLEM, succeeded in suppressing the streak artifacts by iteratively applying a weighting factor; however, the edge of the reconstructed image, which is a high-frequency component in the frequency domain, was enhanced as the number of iterations increased [73]. Thus, the beam profile of the MLEM-reconstructed image increases near the dose fall-off region, as shown in Fig. 14d. The proposed algorithm, FISTA, restored the geometric distortion and decreased the streak artifacts in the reconstructed images while maintaining dose distributions similar to those in the references, as indicated in Fig 14.

The analytic reconstruction algorithm, FBP, severely degraded images with streak artifacts induced by theoretically insufficient angular sampling. The iterative reconstruction algorithm, MLEM, succeeded in suppressing the streak artifacts by iteratively applying a weighting factor; however, the edge of the reconstructed image, which is a high-frequency component in the frequency domain, was enhanced as the number of iterations increased [73]. Thus, the beam profile of the MLEM-reconstructed image increases near the dose fall-off region, as shown in Fig. 14d. The proposed algorithm, FISTA, restored the geometric distortion and decreased the streak artifacts in the reconstructed images while maintaining dose distributions similar to those in the references, as indicated in Fig 14.

2. Characteristics of the proposed dosimeter

The output factor measured by the proposed dosimeter after applying the field size-based correction function is shown in Fig.15a. The output factor measurement was repeated three times using the ionization chamber and developed dosimeter. The measured data are summarized in Table 2. The correction curve based on the computed field size from the extracted mask restored the error of the camera image sensor, similar to that of the ionization chamber (CC13) used as a reference. The masks of the VMAT plans were extracted using the cycleGAN model to avoid incomplete binarization originating from the noise in a single frame (Fig. 16). The predicted masks using UNet and cycleGAN from the CP image in Fig. 3b and the ground truth image generated through DICOM-RT are presented in Fig. 16. The cycleGAN model detected signals from the noises (arrows) more accurately than the UNet (Fig. 16b).

Table 2 Comparisons of the output factor measurements between the ionization chamber (CC13) and proposed dosimeter

Field size [cm2]Ionization chamberProposed dosimeterDeviation (mproposed/mcc13)


Mean (mcc13)Standard deviation (σcc13)Mean (mproposed)Standard deviation (σproposed)
3×30.88230.00010.88970.01821.0084±0.0206
5×50.92880.00000.93650.02071.0083±0.0223
10×101.00000.00011.00000.01931.0000±0.0193
15×151.03740.00101.03910.01521.0016±0.0147
20×201.06360.00001.06030.01360.9969±0.0128
25×251.08190.00001.08480.01841.0027±0.0170
30×301.09660.00011.09110.00810.9950±0.0074

Figure 15.A comparison of the characteristics between the proposed dosimeter, ionization chamber, and EBT3 film analyzed in terms of the (a) output factor, (b) dose linearity, and (c) dose rate response. MU, monitor unit.

Figure 16.(a) A comparison of mask extraction using UNet and proposed algorithm, and (b) enlarged images inside box A. DICOM-RT, digital imaging and communication in medicine of radiotherapy.

Furthermore, the output factor correction function was validated by comparing the dose profiles from the single-segment dose map of MatriXX ResolutionTM and the proposed system (Fig. 17). The single-segment dose map was extracted from the IMRT plan and was selected as the non-uniform radiation field, as shown in Fig. 17a. Although the dose distribution had a heterogenous radiation field, the dose profile of the proposed system was indicated as similar to that of the MatriXX ResolutionTM. The developed system recorded the dose distribution delivered from the medical LINAC with a fixed energy of 6 MV at a consistent sampling rate of the camera. The dose distribution of the single frame was modulated according to the size of the radiation field. Therefore, the equivalent field size-based output correction restored the output difference induced by the intrinsic response of the camera sensor, so that it could be similar to the measurement of the reference.

Figure 17.A comparison of the (a) single-segment dose maps from the intensity-modulated radiotherapy plans and dose profiles measured along EF¯¯ in fields (b) 1, (c) 2, and (d) 3. 2D, two-dimensional.

In the dose profile comparisons, the dose profile of the MatriXX ResolutionTM was smoothed near the edge of the dose map due to the low resolution of the MatriXX ResolutionTM. However, the dose profile of the proposed system showed enhanced sharpness in the high-dose gradient regions.

The linear response of the developed dosimeter was also observed in the dose linearity test (Fig. 15b). The real-time dose distribution measured by the proposed dosimeter at a fixed dose fluctuated due to the aliasing between the beam pulse rate of the LINAC and the sampling rate of the camera. Therefore, we applied an iterative FT-based aliasing correction to the recorded video. The accumulated doses at different dose rates of the proposed dosimeter were within 3% of constancy (Fig. 15c) The maximum difference of the dose-rate dependency in the developed system was 2.35% at 500 MU/min, which is within the daily tolerance recommended by the TG-142 report [10].

3. Dosimetric performances

The gamma analysis of the IMRT plans with gamma criteria of 2%/2-mm and 3%/3-mm are summarized in Table 3. The gamma passing rates for VMAT are listed in Table 4. The gamma passing rate was calculated after specifying the area by a dose threshold of 10%. The average passing rate of the gamma analysis with the 3%/3-mm gamma criterion of the IMRT plans computed between the proposed system and MatriXX ResolutionTM was 98.28%. The mean gamma passing rate using the 3%/3-mm criterion of the IMRT plans computed between the proposed system and radiochromic film was 99.35%. For the VMAT plans, the average passing rate of the gamma analysis with the 3%/3-mm gamma criterion computed between the proposed system and MatriXX ResolutionTM was 98.01%. The mean gamma passing rate using the 3%/3 mm criterion calculated between the proposed system and radiochromic film was 99.27%. The recommended gamma passing rate with the 3%/3-mm gamma criterion was 95% at the clinical sites [9]. The gamma passing rates computed between the proposed system and EBT3 film were higher than those calculated between the proposed system and MatriXX ResolutionTM, since the EBT3 film had a high spatial resolution similar to that of the proposed dosimeter.

Table 3 Quantitative results of the gamma passing rate with the 2%/2-mm and 3%/3-mm gamma criteria for 10 IMRT cases. The gamma analysis was performed between the dose distributions measured from the developed dosimeter and dose map acquired from MatriXX ResolutionTM and radiochromic film

Patient IDSiteTumor volume [cc]MUField/segmentMatriXX ResolutionTMRadiochromic film


2%/2-mm3%/3-mm2%/2-mm3%/3-mm
1Brain6.59604.157/6593.6498.6998.3399.85
2Brain28.02643.585/6688.6396.6690.4498.91
3H&N68.29592.205/4889.3698.6394.3999.45
4H&N19.85882.455/5490.2598.8694.9499.27
5Lung4.871,138.975/4589.7297.9091.7598.31
6Lung79.06605.015/6995.4999.2095.6799.62
7Abdomen123.62918.735/7086.6096.6593.2799.84
8Abdomen351.052,413.065/6092.9899.1595.6999.94
9Pelvis27.09764.916/5792.0898.4895.1099.47
10Pelvis79.15923.927/6693.3498.6093.3598.83
Average gamma passing rate91.2198.2894.2999.35

Table 4 Quantitative results of the gamma passing rate with the 2%/2-mm and 3%/3-mm gamma criteria for 10 VMAT cases. The gamma analysis was performed between the dose distributions measured from the developed dosimeter and the dose map acquired from MatriXX ResolutionTM and radiochromic film

Patient IDSiteTumor volume [cc]MUSequence/CPMatriXX ResolutionTMRadiochromic film


2%/2-mm3%/3-mm2%/2-mm3%/3-mm
1Brain6.59471.952/36094.1798.8596.0099.50
2Brain28.02501.112/18088.1598.4097.8899.89
3H&N68.29428.842/24092.8399.1495.3599.59
4H&N19.85717.182/18090.5798.0498.4799.89
5Lung4.87593.892/36088.7896.6991.2898.59
6Lung79.06538.602/20893.8996.8494.6899.37
7Abdomen123.62341.242/36088.0997.6393.2298.42
8Abdomen351.051,164.232/36089.8198.6993.1699.62
9Pelvis27.09664.323/27091.8598.5893.7098.93
10Pelvis79.15515.392/36087.3097.2391.1598.86
Average gamma passing rate90.5498.0194.4999.27

Furthermore, the gamma analysis on a single segment of the IMRT plan and a single CP of the VMAT plan with the gamma criteria of 2%/2-mm and 3%/3-mm are summarized in Table 5. The average passing rate of the gamma analysis with the 3%/3-mm gamma criterion in the single segment of the IMRT plan was 95.25%. The mean gamma passing rate using the 3%/3-mm criterion in the single CP of the VMAT plan was 80.76%. The single gamma results of the VMAT were less accurate than those of the IMRT due to the noise of the single frame and discrepancy of the sampling rate between the proposed system and MatriXX ResolutionTM.

Table 5 Quantitative results of the gamma passing rate with 2%/2-mm and 3%/3-mm gamma criteria for a single segment of the IMRT cases and single CP of the VMAT cases. Gamma analysis was performed between the dose distributions measured from the developed dosimeter and the dose map acquired from MatriXX ResolutionTM

Patient IDSiteIMRTVMAT


Segment2%/2-mm3%/3-mmCP2%/2-mm3%/3-mm
1Brain6586.0195.176865.9771.43
2Brain6687.9297.863270.8486.28
3H&N4890.3094.924663.5570.53
4H&N5491.0695.134663.2077.67
5Lung4585.6194.604467.5481.59
6Lung6988.5293.963868.5185.93
7Abdomen7086.8594.417862.5284.25
8Abdomen6089.7893.627264.0784.67
9Pelvis5694.6897.705562.5383.12
10Pelvis6681.4695.143667.5582.08
Average gamma passing rate88.2295.2565.6380.76

The dose rates of the LINAC in the pelvic IMRT and VMAT cases are shown in Fig. 18. The dose rates from the measurements estimated the LINAC dose rates with a high accuracy. However, the IMRT case was underestimated in several segments, particularly for small radiation field sizes. The error in the dose rate prediction was the largest when the radiation field size was <1 cm2. The proposed system predicts the LINAC dose rate after applying compensation to the instantaneous doses. The compensation process employed the calculated output factor from the function derived from the computed radiation field sizes and the reference output factors measured by the ionization chamber. The reference output factor measurements were at field sizes of 3×3, 5×5, 10×10, 15×15, 20×20, 25×25, and 30×30 cm2. Therefore, the error in the LINAC dose rate prediction could be attributed to the missing data, such as the output factors of the small radiation field sizes, thereby inducing instability in the compensation function.

Figure 18.Examples of the measured and predicted dose rates from the developed dosimeter: (a) measured dose and estimated equivalent field sizes of the pelvic IMRT case, (b) predicted dose rates of LINAC in the pelvic IMRT case, (c) measured dose and estimated equivalent field sizes in the pelvic VMAT case, and (d) predicted dose rates of LINAC in the pelvic VMAT case. IMRT, intensity-modulated radiotherapy; LINAC, linear accelerator; VMAT, volumetric-modulated arc therapy; MU, monitor unit.

4. Mechanical performances

The MAE calculated from the predicted MLC positions and log files for leaf numbers from 2 to 59 were 0.4800 and 0.3776 mm in banks A and B, respectively. Specifically, the position errors of banks A and B were smaller when the MLC leaves were in the state of retraction from the carriages than in the state of extraction. The MLC position errors investigated for the EBT3 film were 0.4039 and 0.3613 mm for banks A and B, respectively. A tolerance for a leaf positioning accuracy of 1 mm was recommended by TG-142 [10]. The predicted positions of the first and last (60th) leaves were larger than the interleaves owing to the edges of the MLC mask. The MLC mask of the picket fence test was extracted by applying a threshold to the normalized dose image used in the IMRT plans. Thus, the penumbras in the vertical and horizontal directions resulted in the rounded tips of the MLC mask during the segmentation process.

The computed MAEs of the MLC positioning in the clinical cases are listed in Table 6. The mean prediction errors of the MLC position for the IMRT plans with a solid water phantom of 5 cm were 1.7010 and 1.4722 mm for banks A and B, respectively. The average MLC prediction accuracy for the VMAT plans with a 5-cm-thick solid water phantom was 2.8107 mm in bank A and 2.7713 mm in bank B. Although the cameras of the proposed dosimeter captured the dose distribution in the VMAT plan with a frame rate of 16.67 ms, the consecutive motion of the MLC contributed to the degradation of the accuracy in MLC position prediction. The maximum speed of the MLC movements for the VMAT case used in the present study was 27.58 mm s−1 based on the data from the log file, which corresponded to a movement of 0.4668 mm in a single frame.

Table 6 The MAE calculated from the predicted MLC positions and log files for the IMRT and VMAT cases

CaseBankPatient ID

12345678910
IMRTA (mm)1.04731.48931.84611.20761.99722.09382.12292.39521.35171.4592
B (mm)0.95291.19721.43051.24701.59241.56301.85162.15911.19171.5364
VMATA (mm)2.51631.95652.27884.29742.56802.75753.85892.04602.73763.0903
B (mm)2.74331.90082.60424.36622.40392.32584.04001.55702.78692.9853

We developed a scintillator-based multipurpose dosimeter for QA in radiation therapy using an instantaneously interpretable real-time CMOS digital camera. The developed dosimeter measured the planar dose distributions of IMRT and VMAT with a high accuracy and provided an analysis of MLC position tracking with a high sampling frequency. Furthermore, the dose rate could be measured through the readout process of the dose distribution from the proposed dosimeter without the need for additional devices. Specifically, the geometric distortion of the proposed dosimeter system created by the tilted cameras on the corner was calibrated with IRT, which is based on the luminescence intensity function.

The conventional dosimeter, MatriXX ResolutionTM, was applied to the dosimetric evaluation to capture the instant variation of the dose distributions in the IMRT and VMAT plans. The proposed system measured the dose distributions of the IMRT and VMAT plans based on the videos recorded at a frame rate of 16.67 ms, whereas the MatriXX ResolutionTM had the sampling rate of 20 ms. The accumulated dose images of the developed dosimeter were compared with the dose maps measured by conventional dosimeters. The average gamma passing rates were 98.28% and 98.01% in the IMRT and VMAT plans, respectively, when computing the difference between the proposed system and MatriXX ResolutionTM for solid water phantoms of 5 cm with the 3%/3-mm gamma criteria. The average gamma passing rates computed between the proposed system and radiochromic film were 99.35% and 99.27% for the IMRT and VMAT plans, respectively, with the 3%/3-mm gamma criteria. The gamma analysis was performed in a single segment of IMRT and a single CP of VMAT, and the average gamma passing rates were 95.25% and 80.76% for the IMRT and VMAT plans, respectively, with the 3%/3 mm gamma criteria. The dose rates assigned by LINAC were predicted from the sequentially recorded dose images with a high sampling frequency and were verified to be well-matched with the log files.

The accuracy of the MLC position is an important component of VMAT when delivering a prescribed dose to a patient. The MLC positioning can be verified with a high accuracy from the log files generated by the LINAC after beam irradiation. However, the MLC positions in the log files were recorded using an indirect method, such as counting the number of motor rotations for leaf extension. The uncertainties of the log file-based MLC positioning validation were evaluated using various methods [74-78]. To enhance the accuracy and minimize these uncertainties, the developed system employs the direct measurement of MLC positions with high speed and precision. This approach offers real-time validation, potentially reducing errors in treatment delivery and improving the reliability of the mechanical QA of medical LINACs. By implementing this direct measurement technique, the radiation treatment can achieve a more accurate dose distribution, minimize critical organ irradiation, and improve the patient outcomes through a more precise and effective radiotherapy treatments.

The feasibility of the mechanical QA performance in the proposed system was evaluated through the picket fence test recommended in the TG-142 report. Through the picket fence test, the proposed system proved its reliability with acceptable accuracy in the mechanical QA of LINAC. Specifically, the proposed system predicted the MLC position of the VMAT after applying the WaveUNet-based MLC mask extraction algorithm to the single CP dose map. Although the single CP dose map included a high noise due to the intrinsic characteristics of the camera sensor, the WaveUNet-based MLC mask extraction algorithm allowed real-time mechanical QA of LINAC to be performed in the proposed system. However, several clinical cases of the VMAT plans showed real-time MLC positioning errors reaching 4 mm. The large difference between the predicted and reference MLC positions was related to the sampling rate of the camera and the noise in the single-frame dose map. The frame rate of the developed dosimeter was set to 60 FPS; however, a time interval of 16.67 ms was induced in the single frame. The time interval of the single frame resulted in motion blurs in the recorded video when the MLCs were continuously modulated at speeds of up to 27.58 mm s-1 during the beam-on time. Furthermore, in the VMAT case, the signal-to-noise ratio of the single frame was reduced due to the insufficient time of acquiring the signal from the camera sensor. However, it is possible to improve the image quality of the recorded video by using a high-performance recording device. Moreover, the optimization of the MLC mask extraction algorithm, such as WaveUNet in the proposed system can improve the accuracy in MLC position tracking.

A real-time dosimeter using a scintillator plate and a CMOS-based camera was used to analyze the dose rate and MLC positioning error for VMAT [38]. In a published study, the usability of the developed dosimeter was successfully proved by tracking the MLC positions and measuring the dose rates in the VMAT plans with fixed dose rates in sequences between two CPs [79]. In this study, the highly sampled dose maps measured using the developed system allowed for real-time analysis. The dosimeter was validated for the dosimetric and mechanical performances in the IMRT and VMAT plans with a perpendicular composite delivery. Not only the distributions and MLC positions of the static dose rate and radiation field but also those in the dynamic dose rate and radiation field were measured at clinically acceptable levels. Specifically, these dosimetric and mechanical evaluations were performed independently of the machine. TrueBeam STx delivered the prescribed doses in the VMAT plans with consecutively varying dose rates. The proposed system also captured the rapidly changing dose rates and analyzed the dosimetric and mechanical characteristics of the instantaneous dose distribution with a high temporal resolution.

In the present study, we successfully developed an instantaneously interpretable real-time dosimeter for QA in a medical LINAC using a scintillator plate and CMOS-based digital cameras. The dosimeter recorded the scintillation generated by X-rays irradiating the scintillator with a high sampling frequency and converted it into a dose image by applying deep learning-based mask extraction, FT-based dose rate correction, and IRT-based geometric correction. The proposed system fulfills the dosimetric and mechanical performance criteria required for state-of-the-art radiation therapy. The gamma analysis of the 20 clinical cases was clinically acceptable, and the accuracy of MLC positioning met the QA protocol requirement of TG-142.

This work was supported by Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (P0012971).

All relevant data are within the paper and its Supporting Information files.

Conceptualization: Dongyeon Lee, Wonjoong Cheon, Youngyih Han. Data curation: Dongyeon Lee, Sungjin Kim. Formal analysis: Dongyeon Lee, Youngyih Han. Funding acquisition: Youngyih Han. Investigation: Dongyeon Lee, Hyosung Cho, Youngyih Han. Methodology: Dongyeon Lee, Hyosung Cho, Youngyih Han. Project administration: Youngyih Han. Resources: Dongyeon Lee, Sungjin Kim, Youngyih Han. Software: Dongyeon Lee, Sungjin Kim, Youngyih Han. Supervision: Youngyih Han. Validation: Dongyeon Lee, Youngyih Han. Visualization: Dongyeon Lee, Youngyih Han. Writing – original draft: Dongyeon Lee, Youngyih Han. Writing – review & editing: Youngyih Han.

  1. Brahme A. Optimization of stationary and moving beam radiation therapy techniques. Radiother Oncol. 1988;12:129-140.
    Pubmed CrossRef
  2. Otto K. Volumetric modulated arc therapy: IMRT in a single gantry arc. Med Phys. 2008;35:310-317.
    Pubmed CrossRef
  3. Ling CC, Burman C, Chui CS, Kutcher GJ, Leibel SA, LoSasso T, et al. Conformal radiation treatment of prostate cancer using inversely-planned intensity-modulated photon beams produced with dynamic multileaf collimation. Int J Radiat Oncol Biol Phys. 1996;35:721-730.
    Pubmed CrossRef
  4. Bortfeld T, Boyer AL, Schlegel W, Kahler DL, Waldron TJ. Realization and verification of three-dimensional conformal radiotherapy with modulated fields. Int J Radiat Oncol Biol Phys. 1994;30:899-908.
    Pubmed CrossRef
  5. Chui CS, LoSasso T, Spirou S. Dose calculation for photon beams with intensity modulation generated by dynamic jaw or multileaf collimations. Med Phys. 1994;21:1237-1244.
    Pubmed CrossRef
  6. Van Esch A, Bohsung J, Sorvari P, Tenhunen M, Paiusco M, Iori M, et al. Acceptance tests and quality control (QC) procedures for the clinical implementation of intensity modulated radiotherapy (IMRT) using inverse planning and the sliding window technique: experience from five radiotherapy departments. Radiother Oncol. 2002;65:53-70.
    Pubmed CrossRef
  7. Low DA, Moran JM, Dempsey JF, Dong L, Oldham M. Dosimetry tools and techniques for IMRT. Med Phys. 2011;38:1313-1338.
    Pubmed CrossRef
  8. Han Y. Review on the pre-treatment quality assurance for intensity modulated radiation therapy. Prog Med Phys. 2013;24:213-219.
    CrossRef
  9. Ezzell GA, Burmeister JW, Dogan N, LoSasso TJ, Mechalakos JG, Mihailidis D, et al. IMRT commissioning: multiple institution planning and dosimetry comparisons, a report from AAPM Task Group 119. Med Phys. 2009;36:5359-5373.
    Pubmed CrossRef
  10. Klein EE, Hanley J, Bayouth J, Yin FF, Simon W, Dresser S, et al; Task Group 142, American Association of Physicists in Medicine. Task Group 142 report: quality assurance of medical accelerators. Med Phys. 2009;36:4197-4212.
    Pubmed CrossRef
  11. Miften M, Olch A, Mihailidis D, Moran J, Pawlicki T, Molineu A, et al. Tolerance limits and methodologies for IMRT measurement-based verification QA: recommendations of AAPM Task Group No. 218. Med Phys. 2018;45:e53-e83.
    KoreaMed CrossRef
  12. Mehrens H, Taylor P, Followill DS, Kry SF. Survey results of 3D-CRT and IMRT quality assurance practice. J Appl Clin Med Phys. 2020;21:70-76.
    Pubmed KoreaMed CrossRef
  13. Ju SG, Han Y, Kum O, Cheong KH, Shin EH, Shin JS, et al. Comparison of film dosimetry techniques used for quality assurance of intensity modulated radiation therapy. Med Phys. 2010;37:2925-2933.
    Pubmed CrossRef
  14. Sharma M, Singh R, Dutt S, Tomar P, Trivedi G, Robert N. Effect of absorbed dose on post-irradiation coloration and interpretation of polymerization reaction in the Gafchromic EBT3 film. Radiation Physics Chem. 2021;187:109569.
    CrossRef
  15. McCurdy BM, Luchka K, Pistorius S. Dosimetric investigation and portal dose image prediction using an amorphous silicon electronic portal imaging device. Med Phys. 2001;28:911-924.
    Pubmed CrossRef
  16. Van Esch A, Depuydt T, Huyskens DP. The use of an aSi-based EPID for routine absolute dosimetric pre-treatment verification of dynamic IMRT fields. Radiother Oncol. 2004;71:223-234.
    Pubmed CrossRef
  17. Lee C, Menk F, Cadman P, Greer PB. A simple approach to using an amorphous silicon EPID to verify IMRT planar dose maps. Med Phys. 2009;36:984-992.
    Pubmed CrossRef
  18. Warkentin B, Steciw S, Rathee S, Fallone BG. Dosimetric IMRT verification with a flat-panel EPID. Med Phys. 2003;30:3143-3155.
    Pubmed CrossRef
  19. McCurdy BM, Greer PB. Dosimetric properties of an amorphous-silicon EPID used in continuous acquisition mode for application to dynamic and arc IMRT. Med Phys. 2009;36:3028-3039.
    Pubmed CrossRef
  20. Nelms BE, Rasmussen KH, Tome WA. Evaluation of a fast method of EPID-based dosimetry for intensity-modulated radiation therapy. J Appl Clin Med Phys. 2010;11:3185.
    Pubmed KoreaMed CrossRef
  21. Ansbacher W. Three-dimensional portal image-based dose reconstruction in a virtual phantom for rapid evaluation of IMRT plans. Med Phys. 2006;33:3369-3382.
    Pubmed CrossRef
  22. Rowshanfarzad P, Sabet M, O'Connor DJ, Greer PB. Improvement of Varian a-Si EPID dosimetry measurements using a lead-shielded support-arm. Med Dosim. 2012;37:145-151.
    Pubmed CrossRef
  23. Ko L, Kim JO, Siebers JV. Investigation of the optimal backscatter for an aSi electronic portal imaging device. Phys Med Biol. 2004;49:1723-1738.
    Pubmed CrossRef
  24. Rowshanfarzad P, Sabet M, O'Connor DJ, Greer PB. Reduction of the effect of non-uniform backscatter from an E-type support arm of a Varian a-Si EPID used for dosimetry. Phys Med Biol. 2010;55:6617-6632.
    Pubmed CrossRef
  25. Boylan C, McWilliam A, Johnstone E, Rowbottom C. The impact of continuously-variable dose rate VMAT on beam stability, MLC positioning, and overall plan dosimetry. J Appl Clin Med Phys. 2012;13:4023.
    Pubmed KoreaMed CrossRef
  26. Bertelsen A, Lorenzen EL, Brink C. Validation of a new control system for Elekta accelerators facilitating continuously variable dose rate. Med Phys. 2011;38:4802-4810.
    Pubmed CrossRef
  27. Nicolini G, Clivio A, Cozzi L, Fogliata A, Vanetti E. On the impact of dose rate variation upon RapidArc® implementation of volumetric modulated arc therapy. Med Phys. 2011;38:264-271.
    Pubmed CrossRef
  28. Baily NA, Horn RA, Kampp TD. Fluoroscopic visualization of megavoltage therapeutic x ray beams. Int J Radiat Oncol Biol Phys. 1980;6:935-939.
    Pubmed CrossRef
  29. Boon SN, van Luijk P, Schippers JM, Meertens H, Denis JM, Vynckier S, et al. Fast 2D phantom dosimetry for scanning proton beams. Med Phys. 1998;25:464-475.
    Pubmed CrossRef
  30. Ding L, Wu Q, Wang Q, Li Y, Perks RM, Zhao L. Advances on inorganic scintillator-based optic fiber dosimeters. EJNMMI Phys. 2020;7:60.
    Pubmed KoreaMed CrossRef
  31. Goiffon V, Virmontois C, Magnan P, Cervantes P, Corbière F, Estribeau M, et al. Radiation damages in CMOS image sensors: testing and hardening challenges brought by deep sub-micrometer CIS processes. Paper presented at: Proc. SPIE 7826, Sensors, Systems, and Next-Generation Satellites XIV, 78261S. Toulouse, France. 2010.
    CrossRef
  32. Tan J, Büttgen B, Theuwissen AJP. Radiation effects on CMOS image sensors due to X-Rays. Paper presented at: The Eighth International Conference on Advanced Semiconductor Devices and Microsystems. Smolenice, Slovakia: The Eighth International Conference on Advanced Semiconductor Devices and Microsystems. 2010: 279-282.
    KoreaMed CrossRef
  33. Huber A, Sergienko G, Kinna D, Huber V, Milocco A, Mercadier L. Response of the imaging cameras to hard radiation during JET operation. Fusion Eng Des. 2017;123:669-673.
    CrossRef
  34. Boon SN, van Luijk P, Böhringer T, Coray A, Lomax A, Pedroni E, et al. Performance of a fluorescent screen and CCD camera as a two-dimensional dosimetry system for dynamic treatment techniques. Med Phys. 2000;27:2198-2208.
    Pubmed CrossRef
  35. Petric MP, Robar JL, Clark BG. Development and characterization of a tissue equivalent plastic scintillator based dosimetry system. Med Phys. 2006;33:96-105.
    Pubmed CrossRef
  36. Collomb-Patton V, Boher P, Leroux T, Fontbonne JM, Vela A, Batalla A. The DOSIMAP, a high spatial resolution tissue equivalent 2D dosimeter for LINAC QA and IMRT verification. Med Phys. 2009;36:317-328.
    Pubmed CrossRef
  37. Guillot M, Beaulieu L, Archambault L, Beddar S, Gingras L. A new water-equivalent 2D plastic scintillation detectors array for the dosimetry of megavoltage energy photon beams in radiation therapy. Med Phys. 2011;38:6763-6774.
    Pubmed CrossRef
  38. Zin HM, Harris EJ, Osmond JP, Allinson NM, Evans PM. Towards real-time VMAT verification using a prototype, high-speed CMOS active pixel sensor. Phys Med Biol. 2013;58:3359-3375.
    Pubmed CrossRef
  39. Miao X, Sim T. Ambient image recovery and rendering from flash photographs. Paper presented at: IEEE International Conference on Image Processing 2005. Genova, Italy. 2005: II-1038.
    CrossRef
  40. Field R. .
  41. Wueller D. Evaluating digital cameras. Paper presented at: Proc. SPIE 6069, Digital Photography II, 60690K. San Jose, CA, USA. 2006.
    CrossRef
  42. Hsu WF, Chuang KW, Hsu YC. Comparisons of the camera OECF, the ISO speed, and the SFR of digital still-picture cameras. Paper presented at: Proc. SPIE 4080, Input/Output and Imaging Technolgies II. Taipei, Taiwan. 2000.
    CrossRef
  43. Hu Y, Zhu TC. Backscatter correction factor for megavoltage photon beam. Med Phys. 2011;38:5563-5568.
    Pubmed CrossRef
  44. Zhu TC, Ahnesjö A, Lam KL, Li XA, Ma CM, Palta JR, et al; AAPM Therapy Physics Committee Task Group 74. Report of AAPM Therapy Physics Committee Task Group 74: in-air output ratio, Sc, for megavoltage photon beams. Med Phys. 2009;36:5261-5291.
    Pubmed CrossRef
  45. Hesamian MH, Jia W, He X, Kennedy P. Deep learning techniques for medical image segmentation: achievements and challenges. J Digit Imaging. 2019;32:582-596.
    Pubmed KoreaMed CrossRef
  46. Zhou T, Ruan S, Canu S. A review: deep learning for medical image segmentation using multi-modality fusion. Array. 2019;3-4:100004.
    CrossRef
  47. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. Paper presented at: Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. Munich, Germany. 2015.
    CrossRef
  48. Ye JC, Han Y, Cha E. Deep convolutional framelets: a general deep learning framework for inverse problems. SIAM J Imaging Sci. 2018;11:991-1048.
    CrossRef
  49. Han Y, Ye JC. Framing U-Net via deep convolutional framelets: application to sparse-view CT. IEEE Trans Med Imaging. 2018;37:1418-1429.
    Pubmed CrossRef
  50. Ge Y, Wei D, Xue Z, Wang Q, Zhou X, Zhan Y, et al. Unpaired Mr to CT Synthesis with explicit structural constrained adversarial learning. Paper presented at: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). Venice, Italy. 2019: 1096-1099.
    Pubmed CrossRef
  51. Ge Y, Xue Z, Cao T, Liao S. Unpaired whole-body MR to CT synthesis with correlation coefficient constrained adversarial learning. Paper presented at: Proc. SPIE 10949, Medical Imaging 2019: Image Processing, 1094905. San Diego, CA, USA. 2019.
    CrossRef
  52. Lee D, Jeong SW, Kim SJ, Cho H, Park W, Han Y. Improvement of megavoltage computed tomography image quality for adaptive helical tomotherapy using cycleGAN-based image synthesis with small datasets. Med Phys. 2021;48:5593-5610.
    Pubmed CrossRef
  53. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. arXiv. 2015. https://doi.org/10.48550/arXiv.1512.03385.
    Pubmed CrossRef
  54. Isola P, Zhu JY, Zhou T, Efros AA. Image-to-image translation with conditional adversarial networks. Paper presented at: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI, USA. 2017.
    CrossRef
  55. Lasio G, Guerrero M, Goetz W, Lima F, Baulch JE. Effect of varying dose-per-pulse and average dose rate in X-ray beam irradiation on cultured cell survival. Radiat Environ Biophys. 2014;53:671-676.
    Pubmed CrossRef
  56. Posar JA, Davis J, Brace O, Sellin P, Griffith MJ, Dhez O, et al. Characterization of a plastic dosimeter based on organic semiconductor photodiodes and scintillator. Phys Imaging Radiat Oncol. 2020;14:48-52.
    Pubmed KoreaMed CrossRef
  57. Nyquist H. Certain factors affecting telegraph speed. Bell Syst Tech J. 1924;3:324-346.
    CrossRef
  58. Beck A, Teboulle M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J Imaging Sci. 2009;2:183-202.
    CrossRef
  59. Martinet B. [Brève communication. Régularisation d'inéquations variationnelles par approximations successives]. Revue française d'informatique et de recherche opérationnelle. Série rouge. 1970;4:154-158. French.
    CrossRef
  60. Combettes PL, Wajs VR. Signal recovery by proximal forward-backward splitting. Multiscale Model Simul. 2005;4:1168-1200.
    CrossRef
  61. Siddon RL. Fast calculation of the exact radiological path for a three-dimensional CT array. Med Phys. 1985;12:252-255.
    Pubmed CrossRef
  62. Heil CE, Walnut DF. Continuous and discrete wavelet transforms. SIAM Rev. 1989;31:628-666.
    CrossRef
  63. Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv. 2017. https://doi.org/10.48550/arXiv.1412.6980.
  64. Cheon W, Jung H, Lee M, Lee J, Kim SJ, Cho S, et al. Development of a time-resolved mirrorless scintillation detector. PLoS One. 2021;16:e0246742.
    Pubmed KoreaMed CrossRef
  65. Levitan E, Herman GT. A maximum a posteriori probability expectation maximization algorithm for image reconstruction in emission tomography. IEEE Trans Med Imaging. 1987;6:185-192.
    Pubmed CrossRef
  66. Biguri A, Dosanjh M, Hancock S, Soleimani M. TIGRE: a MATLAB-GPU toolbox for CBCT image reconstruction. Biomed Phys Eng Express. 2016;2:055010.
    CrossRef
  67. Low DA, Harms WB, Mutic S, Purdy JA. A technique for the quantitative evaluation of dose distributions. Med Phys. 1998;25:656-661.
    Pubmed CrossRef
  68. Depuydt T, Van Esch A, Huyskens DP. A quantitative evaluation of IMRT dose distributions: refinement and clinical assessment of the gamma evaluation. Radiother Oncol. 2002;62:309-319.
    Pubmed CrossRef
  69. Losasso T. IMRT delivery performance with a varian multileaf collimator. Int J Radiat Oncol Biol Phys. 2008;71:S85-S88.
    Pubmed CrossRef
  70. Li JG, Dempsey JF, Ding L, Liu C, Palta JR. Validation of dynamic MLC-controller log files using a two-dimensional diode array. Med Phys. 2003;30:799-805.
    Pubmed CrossRef
  71. Zygmanski P, Kung JH, Jiang SB, Chin L. Dependence of fluence errors in dynamic IMRT on leaf-positional errors varying with time and leaf number. Med Phys. 2003;30:2736-2749.
    Pubmed CrossRef
  72. Zeidan OA, Li JG, Ranade M, Stell AM, Dempsey JF. Verification of step-and-shoot IMRT delivery using a fast video-based electronic portal imaging device. Med Phys. 2004;31:463-476.
    Pubmed CrossRef
  73. Zeng GL. Maximum-likelihood expectation-maximization algorithm versus windowed filtered backprojection algorithm: a case study. J Nucl Med Technol. 2018;46:129-132.
    Pubmed CrossRef
  74. Hughes J. Automated analysis of Varian log files for advanced radiotherapy treatment verification: a multicenter study [dissertation]. Crawley: University of Western Australia; 2015.
  75. Manikandan A, Sarkar B, Holla R, Vivek TR, Sujatha N. Quality assurance of dynamic parameters in volumetric modulated arc therapy. Br J Radiol. 2012;85:1002-1010.
    Pubmed KoreaMed CrossRef
  76. Rangaraj D, Zhu M, Yang D, Palaniswaamy G, Yaddanapudi S, Wooten OH, et al. Catching errors with patient-specific pretreatment machine log file analysis. Pract Radiat Oncol. 2013;3:80-90.
    Pubmed CrossRef
  77. Agnew A, Agnew CE, Grattan MW, Hounsell AR, McGarry CK. Monitoring daily MLC positional errors using trajectory log files and EPID measurements for IMRT and VMAT deliveries. Phys Med Biol. 2014;59:N49-N63.
    Pubmed CrossRef
  78. Calvo-Ortega JF, Teke T, Moragues S, Pozo M, Casals-Farran J. A Varian DynaLog file-based procedure for patient dose-volume histogram-based IMRT QA. J Appl Clin Med Phys. 2014;15:4665.
    Pubmed KoreaMed CrossRef
  79. Bedford JL, Warrington AP. Commissioning of volumetric modulated arc therapy (VMAT). Int J Radiat Oncol Biol Phys. 2009;73:537-545.
    Pubmed CrossRef

Article

Original Article

Progress in Medical Physics 2024; 35(4): 178-204

Published online December 31, 2024 https://doi.org/10.14316/pmp.2024.35.4.178

Copyright © Korean Society of Medical Physics.

Development of an Instantaneously Interpretable Real-Time Dosimeter System for Quality Assurance of a Medical Linear Accelerator

Dongyeon Lee1,2 , Sung Jin Kim2 , Wonjoong Cheon3 , Hyosung Cho1 , Youngyih Han2,4

1Department of Radiation Convergence Engineering, Yonsei University, Wonju, 2Department of Radiation Oncology, Samsung Medical Center, Seoul, 3Department of Radiation Oncology, Seoul St. Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, 4Department of Health Sciences and Technology, SAIHST, Sungkyunkwan University, Seoul, Korea

Correspondence to:Youngyih Han
(youngyih@skku.edu)
Tel: 82-2-3410-2604
Fax: 82-2-3410-2619

Received: September 23, 2024; Revised: December 13, 2024; Accepted: December 13, 2024

This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Purpose: Modern radiotherapy delivers radiation doses to targets within a few minutes using intricate multiple-beam segments determined with multi-leaf collimators (MLC). Therefore, we propose a scintillator-based dosimetry system capable of assessing the dosimetric and mechanical performance of intensity-modulated radiotherapy (IMRT) and volumetric-modulated arc therapy (VMAT) in real time.
Methods: The dosimeter was equipped with a scintillator plate and two digital cameras. The dose distribution was generated by applying deep learning-based signal processing to correct the intrinsic characteristics of the camera sensor and a tomographic image reconstruction technique to rectify the geometric distortion of the recorded video. Dosimetric evaluations were performed using a gamma analysis against a two-dimensional array and radiochromic film measurements for 20 clinical cases. The average difference in the MLC position measurements and machine log files was tested for the applicability of the mechanical quality assurance (QA) of MLCs.
Results: The agreement of the dose distribution in the IMRT and VMAT plans was clinically acceptable between the proposed system and conventional dosimeters. The average differences in the MLC positions for the IMRT/VMAT plans were 1.7010/2.8107 mm and 1.4722/2.7713 mm in banks A and B, respectively.
Conclusions: In this study, we developed an instantaneously interpretable real-time dosimeter for QA in a medical linear accelerator using a scintillator plate and digital cameras. The feasibility of the proposed system was investigated using dosimetric and mechanical evaluations in the IMRT and VMAT plans. The developed system has clinically acceptable accuracy in both the dosimetric and mechanical QAs of the IMRT and VMAT plans.

Keywords: Dosimeter, Real-time, Deep learning, Dose rate, Multi-leaf collimators

Introduction

Modern radiotherapy techniques allow normal tissues to receive a minimal dose of radiation while maintaining a prescription dose conformal to the target volume. Specifically, intensity-modulated radiotherapy (IMRT) [1] and volumetric-modulated arc therapy (VMAT) [2] deliver radiation doses to a target within a few minutes using intricate multiple beam segments determined with jaw and multi-leaf collimators (MLCs). The combination of multiple beam segments allows the radiation dose to be more focused on the target as compared with the three-dimensional conformal radiotherapy (3DCRT) [3-5]; however, the complexity of dose modulation in IMRT and VMAT requires a more suffocating validation process as compared with 3DCRT [6-8]. Therefore, the task group (TG) of the American Association of Physicists in Medicine (AAPM) has published reports on the patient-specific quality assurance (QA) of IMRT and VMAT plans; these reports have been widely adopted by radiation therapy institutes to reduce the uncertainty of IMRT and VMAT [9-11].

Patient-specific QA can be performed by comparing the measured dose distributions with the dose map computed from the treatment planning system (TPS). According to a recent survey, 80.6% of the sites responded that they conducted patient-specific QA with planar measurements to verify the delivered dose [12]. Planar dosimetry can be performed using a radiochromic film, an electronic portal imaging device (EPID), a two-dimensional (2D) diode array, and a 2D ionization chamber array. Radiochromic films have advantages, including a high spatial resolution and convenience of employment [13]. However, they cannot be reused and are unsuitable for real-time dosimetry because of the saturation time required to stabilize the discoloration [14]. The EPID has shown potential to replace film dosimetry [15,16]. Generally, dosimetry using the EPID was performed by applying the water dose conversion to the acquired radiograph image [17-21], and the properties of the EPID allowed performance of the machine-independent measurements. However, the instability in dosimetry has been presented, since the EPID was originally designed for imaging applications. Furthermore, structural limitations, including increased non-uniform backscatter due to the supporting arm, caused a non-linear response to radiation [22-24].

The continuously varying dose rates of VMAT allows the MLC to be modulated at a high speed such that the delivery time of VMAT is shorter than that of IMRT. Additionally, a continuously varying dose rate contributes to improved beam flatness and stability as compared to a fixed dose rate [25]. However, the high speed of the MLC increases the positioning errors of leaves [26,27]; therefore, the instant dose rate and accuracy of the MLC position are important components of VMAT quality.

A scintillator plate-based dosimeter with a camera was first developed in 1980 by Baily et al. [28] and improved by Boon et al. [29] for the direct measurement of the radiation dose in a water phantom. The characteristics of the scintillator with respect to linearity, repeatability, and stability were evaluated, and its feasibility as a dosimeter was proved [30]. Camera image sensors, including complementary metal-oxide-semiconductors (CMOSs) and charge-coupled devices (CCDs), are damaged when the camera is in the radiation field, which can increase the dark current and hot pixels in the captured image [31-33]. Thus, in the conventional scintillator dosimeter, the camera was mounted on the side and a mirror was installed to protect the image sensor from the primary beam and prevent geometric distortion of the recorded image by reflecting the scintillation [34-37]. However, the built-in mirror increased the dosimeter volume, and the commercial scintillator dosimeter, for example, Lynx PT (IBA Dosimetry) has dimensions of 360×370×600 mm3. Although the accessibility of scintillator dosimeters is limited due to their excessive volume, recent advances in signal processing, which enable cameras to capture images with a high temporal resolution, have shown promise as verification devices for VMAT [38].

In the present study, we developed a mirror-less real-time scintillator dosimeter for the QA of a medical linear accelerator (LINAC), specifically for IMRT and VMAT plans, which can detect the varying dose rates of delivering beams. In this article, the IMRT was defined as a plan with a static MLC during beam delivery, whereas the VMAT was defined as a plan with a dynamic MLC. The dosimeter was composed of a scintillator plate and two CMOS-based cameras without a mirror. The proposed system was capable of measuring the dose distributions at 60 frames per second (FPS), which corresponds to a sampling rate of 16.67 ms. The high temporal resolution provided a function for interpreting the instantaneously modulated dose distribution in the media. Specifically, the proposed system performed a real-time analysis of VMAT, which could predict the dose rates of the LINAC and track the MLC positions in multiple beam segments through highly sampled dose maps. Removing the mirror improved the usability by reducing the weight to 4.08 kg. The characteristics of the developed system were analyzed and the dosimetric and mechanical performance were verified.

Materials and Methods

1. Design of the dosimeter

The proposed dosimeter was composed of a scintillator plate and two digital cameras with CMOS image sensors in an aluminum-frame-based dark box (Fig. 1). A scintillator plate was placed on the top of the dark box, and two digital cameras were installed at the bottom, tilted to face the scintillator plate. Both cameras had the same tilting angle of 57°. The dosimeter had a volume of 316×606×306 mm3 and a weight of 4.08 kg. All sides of the dosimeter were covered with a black paper to block the external light.

Figure 1. Experimental setup of the developed dosimeter with a scintillator plate and complementary metal-oxide-semiconductor based digital cameras.

The PI-200 (Mitsubishi Chemical) set on top of the dark box was a flat-plate-type inorganic scintillator with a size of 300×300×0.9 mm3. It includes three layers of protective phosphor and supporting materials. The phosphor layer was composed of terbium-doped gadolinium oxysulfide, which is widely used in medical imaging systems.

A DSC-RX100M7 (Sony Group Corp.) digital camera recorded the emitted light from the scintillator plate (volume of 42.8×101.6×58.1 mm3 and a weight of 302 g). The recorded video had high-resolution images with pixel dimensions of 1,920×1,080 and a frame rate of 60 FPS, equivalent to 16.67 ms in a single frame. The cameras were operated using remote controllers during irradiation. After recording, the video from the two digital cameras was transmitted to a computer for signal processing, as described below.

2. Signal processing of the proposed system

The light emitted from the scintillator was recorded by a camera using a CMOS image sensor. The main peak in the spectrum of the emitted scintillation was approximately at 530 nm, and the recorded videos are mainly presented in green. Therefore, we performed signal processing using the green channel of the recorded video to convert the recorded video into a dose map. Fig. 2 shows the signal processing of the proposed dosimeter system.

Figure 2. Simplified flow chart of the signal processing in the developed system. Output correction is made in the signal processing. IMRT, intensity-modulated radiotherapy; VMAT, volumetric-modulated arc therapy; FT, Fourier transformed.

First, when the recorded video was input, the video was summed along the frames, and a threshold value was applied to the summed video to obtain a beam signal. Thereafter, the masks of the radiation fields were extracted from the video of the IMRT and VMAT plans through segment and control point (CP)-based processing, respectively. Given that IMRT and VMAT have different mechanisms of dose delivery, we divided the signal processing in two ways. Specifically, the recorded video of the IMRT was disturbed by aliasing between the frame rate of the camera and the fixed beam pulse rate of the LINAC in each segment. Therefore, we applied an iterative Fourier transform (FT)-based aliasing correction to the recorded IMRT video before mask extraction. A detailed description is provided in Section “Iterative Fourier transform-based aliasing correction”. Thereafter, the IMRT mask was generated after applying a threshold of 0.45 to the normalized segment image. The threshold value for the IMRT video mask extraction was selected empirically. The segmented image was acquired by accumulating the video frames recorded at each subfield of the IMRT plan. Fig. 3a shows an example of the recorded video images and the extracted masks from a single segment of the IMRT plan. However, VMAT modulates the radiation doses with consecutively varying radiation fields built from the jaw and MLC, and it is controlled in the LINAC through CPs. The recorded video of the VMAT plan had a different radiation field in each frame corresponding to a CP. However, a single frame of the VMAT plan recording includes grain noises because of the high international organization for standardization (ISO) value setting of the digital cameras [39,40], whereas the recorded video from the IMRT plan was free from grain noises due to the accumulated frames. Consequently, the mask generated by applying a threshold to a single frame of the VMAT plan was incompletely binarized, as shown in Fig. 3b. Therefore, we applied deep learning-based mask extraction to the VMAT plan’s recorded video to overcome this limitation. The details of the deep learning-based mask extraction of the VMAT are described in Section “Video mask extraction of the VMAT plan”.

Figure 3. Extracted mask from the recorded videos of the (a) IMRT and (b) VMAT plans by the thresholding method. The mask of the IMRT plan was extracted from the segmented image accumulated during the beam-on time, whereas that of the VMAT plan was obtained from the single-frame image corresponding to the CP. IMRT, intensity-modulated radiotherapy; VMAT, volumetric-modulated arc therapy; CP, control point.

After the mask extraction, a signal correction was applied to the video image. Camera image sensors, including CMOS and CCD, have intrinsic sensitivity to brightness according to the ISO value. The sensitivity of the camera sensor to a certain ISO value is indicated by an S-shaped curve for the dynamic range of light [41,42]. Generally, camera sensors with higher ISO values respond to photons more quickly as compared to those with lower ISO values. Therefore, cameras in dark environments often use high ISO value settings. In the present study, an ISO value of 12,800 was adopted to record the light of the scintillator plate in a dark box. However, because of the intrinsic response of the camera sensor to brightness, the output factors for several radiation field sizes measured with the proposed dosimeter were different from the reference values measured by the ionization chamber (CC13; IBA Dosimetry), as shown in Fig. 4. The output factor of the proposed dosimeter before the signal correction was observed to increase sharply as compared to that of the reference in large-sized radiation fields.

Figure 4. (a) Comparison of the output factors between the ionization chamber and proposed system before the output correction and (b) correction curve of the proposed dosimeter in the solid water phantom thickness of 5 cm.

A cubic function-based correction curve was employed in the recorded video to mitigate the discrepancy between the camera and ionization chamber measurements. The output factor is dependent on the size of the radiation field; thus, we computed the equivalent field size (Acomputed) from the extracted mask (mask) as follows:

Amask= iX jYmaski,j,
Acomputed=Amask1 +Amask2 2Cmask,

where i and j are the pixel positions of the extracted mask in the vertical and horizontal directions, respectively. Furthermore, A(mask) is the number of pixels selected in the extracted mask from the recorded video, and it was computed from the binarized mask image. The Cmask is the positive constant that converts the pixel size to the equivalent field size. A digital image is presented by numerous pixels of a specific size. The field size of the region of interest can be calculated using the information of the pixel size and number of pixels. Generally, the selected pixel is represented by one and the background is represented by zero in the binary image, such as mask. Therefore, the number of pixels selected in the mask can be computed by adding up the total pixels of the image [Eq. (1)]. However, the recorded videos had geometric distortion due to the position of the digital cameras. In the present study, we used the average number of pixels in two masks (mask1 and mask2) extracted from the recorded videos of two different digital cameras to mitigate the effects of geometric distortion in the images [Eq. (2)]. The equivalent field size was computed after applying the square root to the mean and dividing by the pixel size correction factor Cmask. The linear response between the computed and actual field sizes was observed (Fig. 5).

Figure 5. The linear response between the computed and actual field sizes. The computed square root field size was calculated from the extracted masks of the recorded videos.

Thereafter, the correction curve was estimated using a cubic function based on the curve fitting function of MATLAB (R2021a, MathWorks), and the computed radiation field size from the extracted mask and output factor ratio of the reference to the measurement were used as the input parameters. The output factor correction was performed using the equivalent field size for each frame of the recorded video. The setup had a source-axis distance of 100 cm to a 5-cm deep solid water phantom, where the signal correction curve was estimated. The distributions under different setup conditions from the aforementioned state can be measured after applying signal correction, which is calculated under the setup condition to be observed.

The proposed dosimeter required space under the scintillator plate to capture the light produced as a result of the interaction between the scintillator material and radiation. This structural design makes it difficult to ensure the contribution of the backscatter generated from the materials behind the scintillator plate. Generally, a backscatter thickness of ≥6 cm is required for a 6-MV photon beam [43,44]. However, the signal correction of the proposed system was used to rectify every error caused by the camera sensor and structural limitations, including the backscattering deficiency. After the signal correction, a geometric correction was performed to recover the distortion caused by the position of the digital cameras. The details of the geometric correction are described in Section “Iterative reconstruction technique-based geometric correction”.

The prescribed signal processing procedure for the IMRT and VMAT plans are summarized in Fig. 2.

3. Video mask extraction of the volumetric modulated arc therapy plan

The proposed dosimeter system corrected the output factor errors using the correction curve based on the estimated size of the radiation field from the mask. Thus, it is important to acquire the exact mask of the radiation field from the recorded video images. However, the VMAT plan’s recorded video had a limitation in mask generation using the thresholding method, as described in Section “Video mask extraction of the VMAT plan”. Recently, image processing based on deep learning has demonstrated outstanding performance in the medical field, including image segmentation [45,46]. In particular, the usability of the UNet-based mask extractor has been proven [47]. However, the enlarged receptive field of UNet with pooling layers can lose the details of an image presented at a high frequency [48,49]. Therefore, we applied a cycleGAN-based mask extractor [50-52] to the VMAT plan’s recorded videos.

A simplified diagram of the cycleGAN-based VMAT mask extraction algorithm is shown in Fig. 6. In the proposed algorithm, the cycleGAN model included the following two trainable components: generators and discriminators. The generators and discriminators of cycleGAN work complementarily to achieve convergence during the training process. The generator was trained to produce fake images by reflecting the relationship between the masks (IMask) and video images (IVideo). Specifically, IMask and IVideo were paired inputs and labels, respectively. The discriminator was trained to evaluate whether the images produced by the generator were similar to the label images. In the present study, the generator was composed of nine residual blocks based on the ResNet architecture [53], and the discriminator was adopted from the PatchGAN algorithm [54].

Figure 6. A simplified diagram of mask extraction in the VMAT plan’s recorded video images employing the cycleGAN model. IMRT, intensity-modulated radiotherapy; TPS, treatment planning system; DICOM-RT, digital imaging and communication in medicine of radiotherapy; VMAT, volumetric-modulated arc therapy.

The proposed network was trained with the video images of the IMRT plans and masks were calculated from the digital imaging and communication in medicine of radiotherapy (DICOM-RT) files generated by the TPS (Eclipse ver. 15.5; Varian Medical Systems) at Samsung Medical Center. Generally, the DICOM-RT file includes geometric and dosimetric data specifying the course of radiotherapy, such as the beam angles, collimator openings, and beam modifiers. Therefore, the jaw and MLC positions can be specified from the DICOM-RT file. After extracting the positions of the jaw and MLC, the MLC mask can be generated by calculation using the geometric properties of the medical LINAC. A video mask for training the cycleGAN-based mask extraction algorithm was produced by applying the forward projecting operator (ALumi), described in Section “Iterative reconstruction technique-based geometric correction”, to the MLC mask. The accumulation process for acquiring the segmented images was ignored to generate CP-like images from the IMRT plan’s recorded video.

4. Iterative Fourier transform-based aliasing correction

We used an X-ray beam with an energy of 6 MV from a me&#173;dical LINAC (TrueBeam STx; Varian Medical Systems) at various dose rates. The dose rate was defined as the monitor unit (MU) per minute of the medical LINAC. In general, the dose per pulse of a medical LINAC is constant, and the dose rate is varied by modifying the beam pulse rate [55,56]. The TrueBeam STx used in this study has a beam pulse rate of 60 Hz at a dose rate of 100 MU/min. The beam pulse frequency was sequentially increased to 120 Hz at 200 MU/min, 180 Hz at 300 MU/min, 240 Hz at 400 MU/min, 300 Hz at 500 MU/min, and 360 Hz at 600 MU/min. Therefore, we set the camera frame rate to 60 FPS to cover all configurable dose rates in the TrueBeam STx. The scintillation signal input to the camera can be periodically attenuated owing to the sampling frequency of the image sensor, as shown in Fig. 7a, which is referred to as aliasing [57]. As a result, the signal from a single frame had a linear relationship with the dose rate; however, the dose rate response of the accumulated frames was indicated differently even when the same dose was irradiated. (Fig. 7b) Thus, we applied an iterative FT-based aliasing correction method to the IMRT signals, as shown in Fig. 8. First, the signal from the IMRT was FT to separate the input signal into real and imaginary parts in the frequency domain. The zero-frequency component (FT(x0)) was extracted for use as a criterion for the aliasing correction. Thereafter, the signal (xk) that contributed to building the aliasing in the input signal was selected from the real and imaginary parts as follows:

Figure 7. Comparison of the following signals measured by the proposed dosimeter among the different dose rate settings of the TrueBeam STx (Varian Medical Systems): (a) periodically attenuated signals due to the interference of the sampling frequency between the linear accelerator and cameras, (b) response to the dose rates in the single frame and accumulated frames, and (c) calibrated signal. MU, monitor unit.

Figure 8. Flowchart of the iterative Fourier transform (FT)-based aliasing correction algorithm. DC, direct current.

xk=10k<RealFTxk<10k+1,

where the initial k was set to 1 in. the iteration. The signals selected from the real and imaginary parts were attenuated by 10k to relive the aliasing of the original signal. The real and imaginary parts were successively updated (x*) until k satisfied the following criteria:

k<log10RealFTx0

Finally, the corrected signal was obtained by applying an inverse FT to x*, as shown in Fig. 7c.

Generally, the VMAT delivers the radiation with a continuously variable dose rate during the beam-on time, and the continuously variable dose rate can be indicated as a multi-band signal. The measured signals at different dose rates of the IMRT, including aliasing (Fig. 7a), had low frequencies, and these were negligible at continuously variable beam pulse rates. Therefore, we applied the FT-based aliasing correction step only to the IMRT plans.

5. Iterative reconstruction technique-based geometric correction

The side-positioned cameras caused geometric distortions, including prospective distortions, owing to the tilting angles, as shown in Fig. 9a. Thus, we applied an iterative reconstruction technique (IRT) to the recorded video, allowing geometric distortion to be corrected, even in wide radiation fields. First, we assumed that the proposed dosimeter was a digital tomosynthesis model, as shown in Fig. 9b. Generally, computed tomography (CT) systems can be restored using IRT, a mathematical process that generates tomographic images from the projection data acquired at many different angles with repetitive calculation. In this study, we applied a fast iterative shrinkage-thresholding algorithm (FISTA) [58] to correct the geometric distortion of the proposed dosimeter. The details of the parameters used in the IRT-based geometric correction are listed in Table 1.

Table 1 . Parameters used in the IRT-based geometric correction for dosimetric and mechanical QA.

ParameterDosimetric QAMechanical QA
Source-to-detector distance (mm)543.72543.72
Source-to-object distance (mm)357.71357.71
No. of projections22
Pixel size (mm)0.30.3
Pixel dimension1,920×1,0801,920×1,080
Voxel size (mm)10.1
Voxel dimension300×3003,000×3,000
Reconstruction algorithmFISTAFISTA


Figure 9. (a) Examples of the recorded video with geometric distortion (left) and corresponding dose map corrected with the tomographic image reconstruction technique (right). The geometric distortion of the recorded video was corrected by considering the proposed dosimeter as a (b) digital synthesis model.

The mathematical settings for the description of the FISTA framework in a general CT system were prepared as follows. The original image vector to be reconstructed (x) and the measured projection vector (b) are represented by

x=x1,x2, ...,xNT
b=b1,b2, ...,bMT

where N is the number of voxels, M is the total number of sampling points in the projection data, and the superscript T is the transpose operator. The system matrix is defined as the matrix including the weighting of the interaction model between every image voxel point and all the different rays in the projection:

A=aij,  i=1, 2, ...,M and j=1, 2, ...,N

Using these definitions, the measurement equation is described as follows:

Ax=b

The linear equation Eq. (4) has an infinite solution under the condition M<N, which is physically identical to the CT reconstruction. Thus, the regularization method with l1 norm, which is less sensitive to the sharp edges of the images as compared to l2 regularization, is employed to relieve the instability of Eq. (4) and for optimization. The l1 regularization method enables Eq. (8) to be solved by reducing N to M by exploiting the sparsity of the optimal solution (x*):

x*=minxFxfx+gx
fx=12Axb22
gx=λx1

where f is a smooth convex function and g is a non-smooth convex function owing to the l1 term. The quadratic approximation (QL) is derived from F at a given point y to optimize the objective function by reducing the impact of the non-smooth convex function g. Thereafter, the process of searching the x* as a convex optimization problem (pL) can be rewritten:

QLx,y=fy+xy,fy+L2xy2+gx
pLx,y=argminxQgx+L2xy1Lfy2

where L is a Lipschitz constant and Q is the set of feasible x.

The convex optimization problem described in Eq. (9) can be solved approximately but efficiently using the accelerated proximal gradient descent method [59,60]. In the CT reconstruction process, the system matrix is repeatedly expanded to calculate the forward and backward projections in each iteration loop.

The conventional system matrix (A) for CT reconstruction was computed using Siddon’s algorithm [61]. Siddon’s algorithm calculates the projection data using the intersection length along the voxels of the image volume. However, the proposed dosimeter system measures the emitted light of the scintillator plate, which is recorded using a digital camera following the luminescence intensity function (B):

B=I4πr2

where I is the intensity of the light from the source and r is the distance between the source and measurement point. Therefore, we developed a luminescence intensity-based system matrix (ALumi) to reconstruct the radiation field image and replace A with ALumi in the FISTA framework.

ALumi=14πI1a1,12I2a1,22INa1,N2I1a2,12I2a2,22INa2,N2I1aM,12I2aM,22INaM,N2,

where aM,N is the distance between the center of the image voxel and camera image sensor.

6. Multileaf collimator mask generation

The dose maps measured by the developed dosimeter were built from the jaw and MLC of the LINAC. Traditionally, the radiation field size is defined as the 50% isodose line of the beam profiles at the maximum dose. The MLC positions in the measured dose maps can be predicted using the binarized mask after applying a simple threshold, as shown in Fig. 10.

Figure 10. An example of the multi-leaf collimators (MLC) mask generation in the intensity-modulated radiotherapy (IMRT) plan.

Thresholding-based mask extraction for the prediction of the MLC position in VMAT plans was unable to produce binary images similar to the real MLC position maps because of the increased noise in a single frame, as shown in Fig. 11. Therefore, we applied wavelet-assisted UNet (WaveUNet) to generate MLC masks from the reconstructed CP dose images (Fig. 12).

Figure 11. An example of the multi-leaf collimators (MLC) mask generation in the volumetric-modulated arc therapy (VMAT) plan. CP, control point; WaveUNet, wavelet-assisted UNet.

Figure 12. A simplified diagram representing the network structure of the WaveUNet for extracting the multileaf collimator mask from the control point dose image in the volumetric modulated arc therapy case.

A single scale of the 2D discrete wavelet transform without a diagonal detail [62] was applied to the reconstructed CP dose images before being feedforward to the network. The input wavelet-transformed images were downsized through the pooling layers in the encoding part of the WaveUNet and upsized through the unpooling layers in the decoding part. The WaveUNet includes a shortcut connection indicated by dotted arrows in Fig. 12 to include features of the encoding part in the decoding part, as a conventional UNet model [47]. The WaveUNet model was then optimized using the binary cross-entropy loss function (LBCE) as follows:

LBCE=IlabellogIpred+1Ilabellog1Ipred

where Ilabel represents the paired-label MLC masks computed from the DICOM-RT files and Ipred represents the synthesized binary images when the CP-like dose images from the IMRT plan are entered into WaveUNet. A total of 8,948 paired images were prepared to train the WaveUNet model, and the Adam optimizer [63] optimized the network at a learning rate of 1×10−4 during the training process of 1,000 epochs.

7. Prediction of the linear accelerator dose rate

Given that the developed system measured the absolute dose distributions at a high sampling rate, the dose rate of the LINAC was predictable. The proposed dosimeter measured the dose distribution every 16.67 ms during the beam-on time. First, instantaneous doses were extracted from all the frames recorded during the beam-on time. The LINAC dose rate was then estimated after applying compensation for the measured dose rates. X-rays were delivered from the LINAC during the beam-on time at a constant dose rate; however, the measured dose rates were different because the radiation field sizes and doses were modulated by the jaw and MLCs. Thus, the dose rate of the LINAC was estimated after applying compensation to the instantaneous doses. In the signal correction process, we acquired the reference output factors using the ionization chamber and the equivalent radiation field sizes computed from the extracted video masks. The compensation process was performed by dividing the instantaneous doses by the calculated output factors from the function derived from the computed field sizes and output factors. The predicted LINAC dose rates from the developed dosimeter were compared with those calculated from the TrueBeam trajectory log files.

8. Analysis of the dose map reconstruction quality

In a previous study, an image registration technique was applied to correct the geometric distortion of a scintillator dosimeter using a single camera [64]. The registration-based correction method restored the geometric distortion of the recorded video image by applying the affine transformation, including shear. The registration method was applied to the dual camera-based proposed system to compare the accuracy of the geometric distortion correction with that of the IRT-based method. The IRT-based method used to correct the geometric distortion derived from tilted cameras is widely used in CT reconstruction. Video images from the two cameras were used to reconstruct the dose distribution image of the developed dosimeter, which corresponds to two projection data points in the tomographic system. The other reconstruction algorithms for the analytical and iterative methods were implemented for comparison. Filtered back projection (FBP) and maximum-likelihood estimation–maximization (MLEM) [65] were used as the analytic and iterative methods, respectively, for the comparison of the reconstruction algorithms. The correction of the geometric distortion in the developed dosimeter employing the tomographic image reconstruction technique was performed using a MATLAB code accelerated by a graphic processing unit for fast calculation [66].

9. Validation of the dosimetric performance of the proposed system

The proposed dosimeter system was evaluated in terms of its general, dosimetric, and mechanical performances. The output factor, linearity, and dose-rate dependency were assessed using general performance tests. In the output factor test, we exposed 105 MU equivalent to 100 cGy with field sizes of 3×3, 5×5, 10×10, 15×15, 20×20, 25×25, and 30×30 cm2. The linearity of the developed dosimeter was evaluated using a field size of 10×10 cm2 and delivered doses of 25, 50, 100, 150, 200, and 250 cGy. A fixed dose rate of 400 MU/min was used for the output factor and linearity tests. The reference values for the output factor and linearity test were measured using an ionization chamber (CC13). The dose-rate dependency of the proposed system was observed by irradiating X-ray beams with a dose of 300 cGy and dose rates of 100, 200, 300, 400, 500, and 600 MU/min. In the dose-rate dependency test, the EBT3 film (Ashland) was used as a reference to minimize any influence associated with the sampling frequency issue caused by the digitized dosimeters.

The commercial gamma analysis software (RIT113 ver. 6.2; Radiological Imaging Technology Inc.) was used to evaluate the dosimetric performance of the proposed dosimeter [67,68]. The gamma passing rate was computed to quantify the coincidence between the dose distribution measured by the dosimeter and the dose map acquired from conventional dosimeters, such as a 2D array dosimeter and radiochromic film. The 2D array dosimeter was chosen as it was the sole option capable of instantaneously capturing the modulated dose distributions from the VMAT plans. The IMRT and VMAT plans were generated from 10 clinical cases with five different treatment sites (i.e., brain, head and neck, lung, abdomen, and pelvis); therefore, a total of 20 plans were used in the present study. Given that the proposed system measured the dose distribution by placing the dosimeter on the couch, the clinical plans were recalculated through TPS under the condition of delivering doses from a fixed gantry and collimator set at zero degrees. Specifically, recalculation of the clinical plans to a phantom was performed after anonymizing the individual patient data.

The 2D array dosimeter (MatriXX ResolutionTM; IBA Dosimetry) measured the dose distribution with a pixel spacing of 6.5 mm. The active measurement area of the MatriXX ResolutionTM was 247×247 mm2, and the measured dose map had a resolution of 39×39 pixel. Therefore, we applied linear interpolation to expand the resolution of the MatriXX ResolutionTM dose map to 247×247 with a pixel size of 1×1 mm2. The dose distribution measurement using MatriXX ResolutionTM was performed at a sampling rate of 20 ms to verify the real-time dosimetric performance of the proposed system. For the EBT3 film analysis, digitization was performed with the 75-dots-per-inch scanning option using a 11000XL EPSON flatbed scanner (EPSON). Given that the dose map measured by the proposed dosimeter had a different image size as compared to that of MatriXX ResolutionTM, the dose map of the proposed system was cropped to 247×247. Furthermore, the scanned EBT3 film was rescaled to have a uniform resolution as the dose map of the MatriXX ResolutionTM. Thereafter, co-registration, including image rotation, was applied between the dose map of the proposed system and the measurements of MatriXX ResolutionTM and EBT3 film. Finally, a gamma analysis was performed using the concept of the global gamma through the commercial software RIT113. The real-time gamma analysis for a single CP of the VMAT plan was performed with the extracted CP dose maps, considering the different sampling rates between the developed system and MatriXX ResolutionTM. Specifically, we selected a single CP dose map with a time difference of <0.1 ms between the developed dosimeter and MatriXX ResolutionTM.

10. Validation of the mechanical performance of the proposed system

The quantitative evaluation of the mechanical performance of the developed dosimeter was conducted using the picket fence test [69] recommended by TG-142 to individually validate the alignments of the MLC leaves [10]. The nine narrow bands had a gap width of 10 mm and an interval of 15 mm. TrueBeam STx modulated the radiation fields with 60 pairs of MLC leaves affixed to two opposing parallel carriages (banks A and B). The picket fence test was performed four times with two repetitions each from bank B to bank A and bank A to bank B directions; the test from the bank A to bank B direction was defined as an inverse trial. The MLC positions of the clinical cases were also predicted using the thresholding-based MLC mask in IMRT and the WaveUNet-based MLC mask in VMAT.

The accuracy of MLC positioning was evaluated by comparing the predicted MLC positions in the proposed system and recording the MLC positions in the log files. The log file, which is a record of the LINAC data including the gantry angles, MLC positions, and delivered MU, is stored after treatment completion, and it has been reported to provide accurate MLC positions through various analyses [70-72]. Specifically, the developed dosimeter recorded the dose distribution every 16.67 ms, and the TrueBeam trajectory log file was written with a sampling rate of 20 ms. Therefore, we synchronized the time stamps of the developed dosimeter and log file using linear interpolation to compare the MLC positions in the same phase:

post=posLogt20×20+posLogt20×20posLogt20×20×tt20×2020,

where pos(t) is the MLC position computed from the log file (posLog) at the time of the single-frame dose map in the proposed dosimeter (t). The mechanical performance was validated using the mean absolute error (MAE), which measures the average magnitude of the MLC positioning error between the proposed system and the log file used as a reference:

MAE=i=1npost,ipospredt,in

where i is the leaf number and pospred is the predicted MLC position from the developed dosimeter.

Results

1. Quality of the reconstructed dose map

The examples of the corrected dose distributions obtained by using the registration and IRT-based methods are shown in Figs. 13 and 14. The plots from Figs. 13 and 14 demonstrate the measured beam profiles along AB¯¯ and CD¯¯ as shown in Figs. 13a and 14a, respectively. The registration technique restored the geometric distortion of the proposed dosimeter, similar to the EBT3 film, in a small-sized radiation field. However, the corrected radiation fields were expanded when the size of the radiation fields was increased (Fig 13d), because the registration-based method forcibly stretched the skewed image using the transformation matrix computed during the calibration process [64]. Specifically, the transformation matrix was derived from the four vertices selected by the observer in the checkerboard image within the field of view of the scintillator detector. The IRT-based correction method preserves the reconstruction accuracy of the dose distributions, even for large radiation fields.

Figure 13. A comparison of the (a) correcting methods for geometric distortion and dose profiles in field sizes of (b) 5×5, (c) 10×10, and (d) 20×20 cm2 measured along AB¯¯.

Figure 14. A comparison of the (a) reconstruction algorithms for geometric distortion and dose profiles in field sizes of (b) 5×5, (c) 10×10, and (d) 20×20 cm2 measured along CD¯¯. FBP, filtered back projection; MLEM, maximum-likelihood estimation–maximization.

The analytic reconstruction algorithm, FBP, severely degraded images with streak artifacts induced by theoretically insufficient angular sampling. The iterative reconstruction algorithm, MLEM, succeeded in suppressing the streak artifacts by iteratively applying a weighting factor; however, the edge of the reconstructed image, which is a high-frequency component in the frequency domain, was enhanced as the number of iterations increased [73]. Thus, the beam profile of the MLEM-reconstructed image increases near the dose fall-off region, as shown in Fig. 14d. The proposed algorithm, FISTA, restored the geometric distortion and decreased the streak artifacts in the reconstructed images while maintaining dose distributions similar to those in the references, as indicated in Fig 14.

The analytic reconstruction algorithm, FBP, severely degraded images with streak artifacts induced by theoretically insufficient angular sampling. The iterative reconstruction algorithm, MLEM, succeeded in suppressing the streak artifacts by iteratively applying a weighting factor; however, the edge of the reconstructed image, which is a high-frequency component in the frequency domain, was enhanced as the number of iterations increased [73]. Thus, the beam profile of the MLEM-reconstructed image increases near the dose fall-off region, as shown in Fig. 14d. The proposed algorithm, FISTA, restored the geometric distortion and decreased the streak artifacts in the reconstructed images while maintaining dose distributions similar to those in the references, as indicated in Fig 14.

2. Characteristics of the proposed dosimeter

The output factor measured by the proposed dosimeter after applying the field size-based correction function is shown in Fig.15a. The output factor measurement was repeated three times using the ionization chamber and developed dosimeter. The measured data are summarized in Table 2. The correction curve based on the computed field size from the extracted mask restored the error of the camera image sensor, similar to that of the ionization chamber (CC13) used as a reference. The masks of the VMAT plans were extracted using the cycleGAN model to avoid incomplete binarization originating from the noise in a single frame (Fig. 16). The predicted masks using UNet and cycleGAN from the CP image in Fig. 3b and the ground truth image generated through DICOM-RT are presented in Fig. 16. The cycleGAN model detected signals from the noises (arrows) more accurately than the UNet (Fig. 16b).

Table 2 . Comparisons of the output factor measurements between the ionization chamber (CC13) and proposed dosimeter.

Field size [cm2]Ionization chamberProposed dosimeterDeviation (mproposed/mcc13)


Mean (mcc13)Standard deviation (σcc13)Mean (mproposed)Standard deviation (σproposed)
3×30.88230.00010.88970.01821.0084±0.0206
5×50.92880.00000.93650.02071.0083±0.0223
10×101.00000.00011.00000.01931.0000±0.0193
15×151.03740.00101.03910.01521.0016±0.0147
20×201.06360.00001.06030.01360.9969±0.0128
25×251.08190.00001.08480.01841.0027±0.0170
30×301.09660.00011.09110.00810.9950±0.0074


Figure 15. A comparison of the characteristics between the proposed dosimeter, ionization chamber, and EBT3 film analyzed in terms of the (a) output factor, (b) dose linearity, and (c) dose rate response. MU, monitor unit.

Figure 16. (a) A comparison of mask extraction using UNet and proposed algorithm, and (b) enlarged images inside box A. DICOM-RT, digital imaging and communication in medicine of radiotherapy.

Furthermore, the output factor correction function was validated by comparing the dose profiles from the single-segment dose map of MatriXX ResolutionTM and the proposed system (Fig. 17). The single-segment dose map was extracted from the IMRT plan and was selected as the non-uniform radiation field, as shown in Fig. 17a. Although the dose distribution had a heterogenous radiation field, the dose profile of the proposed system was indicated as similar to that of the MatriXX ResolutionTM. The developed system recorded the dose distribution delivered from the medical LINAC with a fixed energy of 6 MV at a consistent sampling rate of the camera. The dose distribution of the single frame was modulated according to the size of the radiation field. Therefore, the equivalent field size-based output correction restored the output difference induced by the intrinsic response of the camera sensor, so that it could be similar to the measurement of the reference.

Figure 17. A comparison of the (a) single-segment dose maps from the intensity-modulated radiotherapy plans and dose profiles measured along EF¯¯ in fields (b) 1, (c) 2, and (d) 3. 2D, two-dimensional.

In the dose profile comparisons, the dose profile of the MatriXX ResolutionTM was smoothed near the edge of the dose map due to the low resolution of the MatriXX ResolutionTM. However, the dose profile of the proposed system showed enhanced sharpness in the high-dose gradient regions.

The linear response of the developed dosimeter was also observed in the dose linearity test (Fig. 15b). The real-time dose distribution measured by the proposed dosimeter at a fixed dose fluctuated due to the aliasing between the beam pulse rate of the LINAC and the sampling rate of the camera. Therefore, we applied an iterative FT-based aliasing correction to the recorded video. The accumulated doses at different dose rates of the proposed dosimeter were within 3% of constancy (Fig. 15c) The maximum difference of the dose-rate dependency in the developed system was 2.35% at 500 MU/min, which is within the daily tolerance recommended by the TG-142 report [10].

3. Dosimetric performances

The gamma analysis of the IMRT plans with gamma criteria of 2%/2-mm and 3%/3-mm are summarized in Table 3. The gamma passing rates for VMAT are listed in Table 4. The gamma passing rate was calculated after specifying the area by a dose threshold of 10%. The average passing rate of the gamma analysis with the 3%/3-mm gamma criterion of the IMRT plans computed between the proposed system and MatriXX ResolutionTM was 98.28%. The mean gamma passing rate using the 3%/3-mm criterion of the IMRT plans computed between the proposed system and radiochromic film was 99.35%. For the VMAT plans, the average passing rate of the gamma analysis with the 3%/3-mm gamma criterion computed between the proposed system and MatriXX ResolutionTM was 98.01%. The mean gamma passing rate using the 3%/3 mm criterion calculated between the proposed system and radiochromic film was 99.27%. The recommended gamma passing rate with the 3%/3-mm gamma criterion was 95% at the clinical sites [9]. The gamma passing rates computed between the proposed system and EBT3 film were higher than those calculated between the proposed system and MatriXX ResolutionTM, since the EBT3 film had a high spatial resolution similar to that of the proposed dosimeter.

Table 3 . Quantitative results of the gamma passing rate with the 2%/2-mm and 3%/3-mm gamma criteria for 10 IMRT cases. The gamma analysis was performed between the dose distributions measured from the developed dosimeter and dose map acquired from MatriXX ResolutionTM and radiochromic film.

Patient IDSiteTumor volume [cc]MUField/segmentMatriXX ResolutionTMRadiochromic film


2%/2-mm3%/3-mm2%/2-mm3%/3-mm
1Brain6.59604.157/6593.6498.6998.3399.85
2Brain28.02643.585/6688.6396.6690.4498.91
3H&N68.29592.205/4889.3698.6394.3999.45
4H&N19.85882.455/5490.2598.8694.9499.27
5Lung4.871,138.975/4589.7297.9091.7598.31
6Lung79.06605.015/6995.4999.2095.6799.62
7Abdomen123.62918.735/7086.6096.6593.2799.84
8Abdomen351.052,413.065/6092.9899.1595.6999.94
9Pelvis27.09764.916/5792.0898.4895.1099.47
10Pelvis79.15923.927/6693.3498.6093.3598.83
Average gamma passing rate91.2198.2894.2999.35


Table 4 . Quantitative results of the gamma passing rate with the 2%/2-mm and 3%/3-mm gamma criteria for 10 VMAT cases. The gamma analysis was performed between the dose distributions measured from the developed dosimeter and the dose map acquired from MatriXX ResolutionTM and radiochromic film.

Patient IDSiteTumor volume [cc]MUSequence/CPMatriXX ResolutionTMRadiochromic film


2%/2-mm3%/3-mm2%/2-mm3%/3-mm
1Brain6.59471.952/36094.1798.8596.0099.50
2Brain28.02501.112/18088.1598.4097.8899.89
3H&N68.29428.842/24092.8399.1495.3599.59
4H&N19.85717.182/18090.5798.0498.4799.89
5Lung4.87593.892/36088.7896.6991.2898.59
6Lung79.06538.602/20893.8996.8494.6899.37
7Abdomen123.62341.242/36088.0997.6393.2298.42
8Abdomen351.051,164.232/36089.8198.6993.1699.62
9Pelvis27.09664.323/27091.8598.5893.7098.93
10Pelvis79.15515.392/36087.3097.2391.1598.86
Average gamma passing rate90.5498.0194.4999.27


Furthermore, the gamma analysis on a single segment of the IMRT plan and a single CP of the VMAT plan with the gamma criteria of 2%/2-mm and 3%/3-mm are summarized in Table 5. The average passing rate of the gamma analysis with the 3%/3-mm gamma criterion in the single segment of the IMRT plan was 95.25%. The mean gamma passing rate using the 3%/3-mm criterion in the single CP of the VMAT plan was 80.76%. The single gamma results of the VMAT were less accurate than those of the IMRT due to the noise of the single frame and discrepancy of the sampling rate between the proposed system and MatriXX ResolutionTM.

Table 5 . Quantitative results of the gamma passing rate with 2%/2-mm and 3%/3-mm gamma criteria for a single segment of the IMRT cases and single CP of the VMAT cases. Gamma analysis was performed between the dose distributions measured from the developed dosimeter and the dose map acquired from MatriXX ResolutionTM.

Patient IDSiteIMRTVMAT


Segment2%/2-mm3%/3-mmCP2%/2-mm3%/3-mm
1Brain6586.0195.176865.9771.43
2Brain6687.9297.863270.8486.28
3H&N4890.3094.924663.5570.53
4H&N5491.0695.134663.2077.67
5Lung4585.6194.604467.5481.59
6Lung6988.5293.963868.5185.93
7Abdomen7086.8594.417862.5284.25
8Abdomen6089.7893.627264.0784.67
9Pelvis5694.6897.705562.5383.12
10Pelvis6681.4695.143667.5582.08
Average gamma passing rate88.2295.2565.6380.76


The dose rates of the LINAC in the pelvic IMRT and VMAT cases are shown in Fig. 18. The dose rates from the measurements estimated the LINAC dose rates with a high accuracy. However, the IMRT case was underestimated in several segments, particularly for small radiation field sizes. The error in the dose rate prediction was the largest when the radiation field size was <1 cm2. The proposed system predicts the LINAC dose rate after applying compensation to the instantaneous doses. The compensation process employed the calculated output factor from the function derived from the computed radiation field sizes and the reference output factors measured by the ionization chamber. The reference output factor measurements were at field sizes of 3×3, 5×5, 10×10, 15×15, 20×20, 25×25, and 30×30 cm2. Therefore, the error in the LINAC dose rate prediction could be attributed to the missing data, such as the output factors of the small radiation field sizes, thereby inducing instability in the compensation function.

Figure 18. Examples of the measured and predicted dose rates from the developed dosimeter: (a) measured dose and estimated equivalent field sizes of the pelvic IMRT case, (b) predicted dose rates of LINAC in the pelvic IMRT case, (c) measured dose and estimated equivalent field sizes in the pelvic VMAT case, and (d) predicted dose rates of LINAC in the pelvic VMAT case. IMRT, intensity-modulated radiotherapy; LINAC, linear accelerator; VMAT, volumetric-modulated arc therapy; MU, monitor unit.

4. Mechanical performances

The MAE calculated from the predicted MLC positions and log files for leaf numbers from 2 to 59 were 0.4800 and 0.3776 mm in banks A and B, respectively. Specifically, the position errors of banks A and B were smaller when the MLC leaves were in the state of retraction from the carriages than in the state of extraction. The MLC position errors investigated for the EBT3 film were 0.4039 and 0.3613 mm for banks A and B, respectively. A tolerance for a leaf positioning accuracy of 1 mm was recommended by TG-142 [10]. The predicted positions of the first and last (60th) leaves were larger than the interleaves owing to the edges of the MLC mask. The MLC mask of the picket fence test was extracted by applying a threshold to the normalized dose image used in the IMRT plans. Thus, the penumbras in the vertical and horizontal directions resulted in the rounded tips of the MLC mask during the segmentation process.

The computed MAEs of the MLC positioning in the clinical cases are listed in Table 6. The mean prediction errors of the MLC position for the IMRT plans with a solid water phantom of 5 cm were 1.7010 and 1.4722 mm for banks A and B, respectively. The average MLC prediction accuracy for the VMAT plans with a 5-cm-thick solid water phantom was 2.8107 mm in bank A and 2.7713 mm in bank B. Although the cameras of the proposed dosimeter captured the dose distribution in the VMAT plan with a frame rate of 16.67 ms, the consecutive motion of the MLC contributed to the degradation of the accuracy in MLC position prediction. The maximum speed of the MLC movements for the VMAT case used in the present study was 27.58 mm s−1 based on the data from the log file, which corresponded to a movement of 0.4668 mm in a single frame.

Table 6 . The MAE calculated from the predicted MLC positions and log files for the IMRT and VMAT cases.

CaseBankPatient ID

12345678910
IMRTA (mm)1.04731.48931.84611.20761.99722.09382.12292.39521.35171.4592
B (mm)0.95291.19721.43051.24701.59241.56301.85162.15911.19171.5364
VMATA (mm)2.51631.95652.27884.29742.56802.75753.85892.04602.73763.0903
B (mm)2.74331.90082.60424.36622.40392.32584.04001.55702.78692.9853

Discussion

We developed a scintillator-based multipurpose dosimeter for QA in radiation therapy using an instantaneously interpretable real-time CMOS digital camera. The developed dosimeter measured the planar dose distributions of IMRT and VMAT with a high accuracy and provided an analysis of MLC position tracking with a high sampling frequency. Furthermore, the dose rate could be measured through the readout process of the dose distribution from the proposed dosimeter without the need for additional devices. Specifically, the geometric distortion of the proposed dosimeter system created by the tilted cameras on the corner was calibrated with IRT, which is based on the luminescence intensity function.

The conventional dosimeter, MatriXX ResolutionTM, was applied to the dosimetric evaluation to capture the instant variation of the dose distributions in the IMRT and VMAT plans. The proposed system measured the dose distributions of the IMRT and VMAT plans based on the videos recorded at a frame rate of 16.67 ms, whereas the MatriXX ResolutionTM had the sampling rate of 20 ms. The accumulated dose images of the developed dosimeter were compared with the dose maps measured by conventional dosimeters. The average gamma passing rates were 98.28% and 98.01% in the IMRT and VMAT plans, respectively, when computing the difference between the proposed system and MatriXX ResolutionTM for solid water phantoms of 5 cm with the 3%/3-mm gamma criteria. The average gamma passing rates computed between the proposed system and radiochromic film were 99.35% and 99.27% for the IMRT and VMAT plans, respectively, with the 3%/3-mm gamma criteria. The gamma analysis was performed in a single segment of IMRT and a single CP of VMAT, and the average gamma passing rates were 95.25% and 80.76% for the IMRT and VMAT plans, respectively, with the 3%/3 mm gamma criteria. The dose rates assigned by LINAC were predicted from the sequentially recorded dose images with a high sampling frequency and were verified to be well-matched with the log files.

The accuracy of the MLC position is an important component of VMAT when delivering a prescribed dose to a patient. The MLC positioning can be verified with a high accuracy from the log files generated by the LINAC after beam irradiation. However, the MLC positions in the log files were recorded using an indirect method, such as counting the number of motor rotations for leaf extension. The uncertainties of the log file-based MLC positioning validation were evaluated using various methods [74-78]. To enhance the accuracy and minimize these uncertainties, the developed system employs the direct measurement of MLC positions with high speed and precision. This approach offers real-time validation, potentially reducing errors in treatment delivery and improving the reliability of the mechanical QA of medical LINACs. By implementing this direct measurement technique, the radiation treatment can achieve a more accurate dose distribution, minimize critical organ irradiation, and improve the patient outcomes through a more precise and effective radiotherapy treatments.

The feasibility of the mechanical QA performance in the proposed system was evaluated through the picket fence test recommended in the TG-142 report. Through the picket fence test, the proposed system proved its reliability with acceptable accuracy in the mechanical QA of LINAC. Specifically, the proposed system predicted the MLC position of the VMAT after applying the WaveUNet-based MLC mask extraction algorithm to the single CP dose map. Although the single CP dose map included a high noise due to the intrinsic characteristics of the camera sensor, the WaveUNet-based MLC mask extraction algorithm allowed real-time mechanical QA of LINAC to be performed in the proposed system. However, several clinical cases of the VMAT plans showed real-time MLC positioning errors reaching 4 mm. The large difference between the predicted and reference MLC positions was related to the sampling rate of the camera and the noise in the single-frame dose map. The frame rate of the developed dosimeter was set to 60 FPS; however, a time interval of 16.67 ms was induced in the single frame. The time interval of the single frame resulted in motion blurs in the recorded video when the MLCs were continuously modulated at speeds of up to 27.58 mm s-1 during the beam-on time. Furthermore, in the VMAT case, the signal-to-noise ratio of the single frame was reduced due to the insufficient time of acquiring the signal from the camera sensor. However, it is possible to improve the image quality of the recorded video by using a high-performance recording device. Moreover, the optimization of the MLC mask extraction algorithm, such as WaveUNet in the proposed system can improve the accuracy in MLC position tracking.

A real-time dosimeter using a scintillator plate and a CMOS-based camera was used to analyze the dose rate and MLC positioning error for VMAT [38]. In a published study, the usability of the developed dosimeter was successfully proved by tracking the MLC positions and measuring the dose rates in the VMAT plans with fixed dose rates in sequences between two CPs [79]. In this study, the highly sampled dose maps measured using the developed system allowed for real-time analysis. The dosimeter was validated for the dosimetric and mechanical performances in the IMRT and VMAT plans with a perpendicular composite delivery. Not only the distributions and MLC positions of the static dose rate and radiation field but also those in the dynamic dose rate and radiation field were measured at clinically acceptable levels. Specifically, these dosimetric and mechanical evaluations were performed independently of the machine. TrueBeam STx delivered the prescribed doses in the VMAT plans with consecutively varying dose rates. The proposed system also captured the rapidly changing dose rates and analyzed the dosimetric and mechanical characteristics of the instantaneous dose distribution with a high temporal resolution.

Conclusions

In the present study, we successfully developed an instantaneously interpretable real-time dosimeter for QA in a medical LINAC using a scintillator plate and CMOS-based digital cameras. The dosimeter recorded the scintillation generated by X-rays irradiating the scintillator with a high sampling frequency and converted it into a dose image by applying deep learning-based mask extraction, FT-based dose rate correction, and IRT-based geometric correction. The proposed system fulfills the dosimetric and mechanical performance criteria required for state-of-the-art radiation therapy. The gamma analysis of the 20 clinical cases was clinically acceptable, and the accuracy of MLC positioning met the QA protocol requirement of TG-142.

Funding

This work was supported by Korea Institute for Advancement of Technology (KIAT) grant funded by the Korea Government (MOTIE) (P0012971).

Conflicts of Interest

The authors have nothing to disclose.

Availability of Data and Materials

All relevant data are within the paper and its Supporting Information files.

Author Contributions

Conceptualization: Dongyeon Lee, Wonjoong Cheon, Youngyih Han. Data curation: Dongyeon Lee, Sungjin Kim. Formal analysis: Dongyeon Lee, Youngyih Han. Funding acquisition: Youngyih Han. Investigation: Dongyeon Lee, Hyosung Cho, Youngyih Han. Methodology: Dongyeon Lee, Hyosung Cho, Youngyih Han. Project administration: Youngyih Han. Resources: Dongyeon Lee, Sungjin Kim, Youngyih Han. Software: Dongyeon Lee, Sungjin Kim, Youngyih Han. Supervision: Youngyih Han. Validation: Dongyeon Lee, Youngyih Han. Visualization: Dongyeon Lee, Youngyih Han. Writing – original draft: Dongyeon Lee, Youngyih Han. Writing – review & editing: Youngyih Han.

Fig 1.

Figure 1.Experimental setup of the developed dosimeter with a scintillator plate and complementary metal-oxide-semiconductor based digital cameras.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 2.

Figure 2.Simplified flow chart of the signal processing in the developed system. Output correction is made in the signal processing. IMRT, intensity-modulated radiotherapy; VMAT, volumetric-modulated arc therapy; FT, Fourier transformed.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 3.

Figure 3.Extracted mask from the recorded videos of the (a) IMRT and (b) VMAT plans by the thresholding method. The mask of the IMRT plan was extracted from the segmented image accumulated during the beam-on time, whereas that of the VMAT plan was obtained from the single-frame image corresponding to the CP. IMRT, intensity-modulated radiotherapy; VMAT, volumetric-modulated arc therapy; CP, control point.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 4.

Figure 4.(a) Comparison of the output factors between the ionization chamber and proposed system before the output correction and (b) correction curve of the proposed dosimeter in the solid water phantom thickness of 5 cm.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 5.

Figure 5.The linear response between the computed and actual field sizes. The computed square root field size was calculated from the extracted masks of the recorded videos.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 6.

Figure 6.A simplified diagram of mask extraction in the VMAT plan’s recorded video images employing the cycleGAN model. IMRT, intensity-modulated radiotherapy; TPS, treatment planning system; DICOM-RT, digital imaging and communication in medicine of radiotherapy; VMAT, volumetric-modulated arc therapy.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 7.

Figure 7.Comparison of the following signals measured by the proposed dosimeter among the different dose rate settings of the TrueBeam STx (Varian Medical Systems): (a) periodically attenuated signals due to the interference of the sampling frequency between the linear accelerator and cameras, (b) response to the dose rates in the single frame and accumulated frames, and (c) calibrated signal. MU, monitor unit.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 8.

Figure 8.Flowchart of the iterative Fourier transform (FT)-based aliasing correction algorithm. DC, direct current.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 9.

Figure 9.(a) Examples of the recorded video with geometric distortion (left) and corresponding dose map corrected with the tomographic image reconstruction technique (right). The geometric distortion of the recorded video was corrected by considering the proposed dosimeter as a (b) digital synthesis model.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 10.

Figure 10.An example of the multi-leaf collimators (MLC) mask generation in the intensity-modulated radiotherapy (IMRT) plan.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 11.

Figure 11.An example of the multi-leaf collimators (MLC) mask generation in the volumetric-modulated arc therapy (VMAT) plan. CP, control point; WaveUNet, wavelet-assisted UNet.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 12.

Figure 12.A simplified diagram representing the network structure of the WaveUNet for extracting the multileaf collimator mask from the control point dose image in the volumetric modulated arc therapy case.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 13.

Figure 13.A comparison of the (a) correcting methods for geometric distortion and dose profiles in field sizes of (b) 5×5, (c) 10×10, and (d) 20×20 cm2 measured along AB¯¯.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 14.

Figure 14.A comparison of the (a) reconstruction algorithms for geometric distortion and dose profiles in field sizes of (b) 5×5, (c) 10×10, and (d) 20×20 cm2 measured along CD¯¯. FBP, filtered back projection; MLEM, maximum-likelihood estimation–maximization.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 15.

Figure 15.A comparison of the characteristics between the proposed dosimeter, ionization chamber, and EBT3 film analyzed in terms of the (a) output factor, (b) dose linearity, and (c) dose rate response. MU, monitor unit.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 16.

Figure 16.(a) A comparison of mask extraction using UNet and proposed algorithm, and (b) enlarged images inside box A. DICOM-RT, digital imaging and communication in medicine of radiotherapy.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 17.

Figure 17.A comparison of the (a) single-segment dose maps from the intensity-modulated radiotherapy plans and dose profiles measured along EF¯¯ in fields (b) 1, (c) 2, and (d) 3. 2D, two-dimensional.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Fig 18.

Figure 18.Examples of the measured and predicted dose rates from the developed dosimeter: (a) measured dose and estimated equivalent field sizes of the pelvic IMRT case, (b) predicted dose rates of LINAC in the pelvic IMRT case, (c) measured dose and estimated equivalent field sizes in the pelvic VMAT case, and (d) predicted dose rates of LINAC in the pelvic VMAT case. IMRT, intensity-modulated radiotherapy; LINAC, linear accelerator; VMAT, volumetric-modulated arc therapy; MU, monitor unit.
Progress in Medical Physics 2024; 35: 178-204https://doi.org/10.14316/pmp.2024.35.4.178

Table 1 Parameters used in the IRT-based geometric correction for dosimetric and mechanical QA

ParameterDosimetric QAMechanical QA
Source-to-detector distance (mm)543.72543.72
Source-to-object distance (mm)357.71357.71
No. of projections22
Pixel size (mm)0.30.3
Pixel dimension1,920×1,0801,920×1,080
Voxel size (mm)10.1
Voxel dimension300×3003,000×3,000
Reconstruction algorithmFISTAFISTA

Table 2 Comparisons of the output factor measurements between the ionization chamber (CC13) and proposed dosimeter

Field size [cm2]Ionization chamberProposed dosimeterDeviation (mproposed/mcc13)


Mean (mcc13)Standard deviation (σcc13)Mean (mproposed)Standard deviation (σproposed)
3×30.88230.00010.88970.01821.0084±0.0206
5×50.92880.00000.93650.02071.0083±0.0223
10×101.00000.00011.00000.01931.0000±0.0193
15×151.03740.00101.03910.01521.0016±0.0147
20×201.06360.00001.06030.01360.9969±0.0128
25×251.08190.00001.08480.01841.0027±0.0170
30×301.09660.00011.09110.00810.9950±0.0074

Table 3 Quantitative results of the gamma passing rate with the 2%/2-mm and 3%/3-mm gamma criteria for 10 IMRT cases. The gamma analysis was performed between the dose distributions measured from the developed dosimeter and dose map acquired from MatriXX ResolutionTM and radiochromic film

Patient IDSiteTumor volume [cc]MUField/segmentMatriXX ResolutionTMRadiochromic film


2%/2-mm3%/3-mm2%/2-mm3%/3-mm
1Brain6.59604.157/6593.6498.6998.3399.85
2Brain28.02643.585/6688.6396.6690.4498.91
3H&N68.29592.205/4889.3698.6394.3999.45
4H&N19.85882.455/5490.2598.8694.9499.27
5Lung4.871,138.975/4589.7297.9091.7598.31
6Lung79.06605.015/6995.4999.2095.6799.62
7Abdomen123.62918.735/7086.6096.6593.2799.84
8Abdomen351.052,413.065/6092.9899.1595.6999.94
9Pelvis27.09764.916/5792.0898.4895.1099.47
10Pelvis79.15923.927/6693.3498.6093.3598.83
Average gamma passing rate91.2198.2894.2999.35

Table 4 Quantitative results of the gamma passing rate with the 2%/2-mm and 3%/3-mm gamma criteria for 10 VMAT cases. The gamma analysis was performed between the dose distributions measured from the developed dosimeter and the dose map acquired from MatriXX ResolutionTM and radiochromic film

Patient IDSiteTumor volume [cc]MUSequence/CPMatriXX ResolutionTMRadiochromic film


2%/2-mm3%/3-mm2%/2-mm3%/3-mm
1Brain6.59471.952/36094.1798.8596.0099.50
2Brain28.02501.112/18088.1598.4097.8899.89
3H&N68.29428.842/24092.8399.1495.3599.59
4H&N19.85717.182/18090.5798.0498.4799.89
5Lung4.87593.892/36088.7896.6991.2898.59
6Lung79.06538.602/20893.8996.8494.6899.37
7Abdomen123.62341.242/36088.0997.6393.2298.42
8Abdomen351.051,164.232/36089.8198.6993.1699.62
9Pelvis27.09664.323/27091.8598.5893.7098.93
10Pelvis79.15515.392/36087.3097.2391.1598.86
Average gamma passing rate90.5498.0194.4999.27

Table 5 Quantitative results of the gamma passing rate with 2%/2-mm and 3%/3-mm gamma criteria for a single segment of the IMRT cases and single CP of the VMAT cases. Gamma analysis was performed between the dose distributions measured from the developed dosimeter and the dose map acquired from MatriXX ResolutionTM

Patient IDSiteIMRTVMAT


Segment2%/2-mm3%/3-mmCP2%/2-mm3%/3-mm
1Brain6586.0195.176865.9771.43
2Brain6687.9297.863270.8486.28
3H&N4890.3094.924663.5570.53
4H&N5491.0695.134663.2077.67
5Lung4585.6194.604467.5481.59
6Lung6988.5293.963868.5185.93
7Abdomen7086.8594.417862.5284.25
8Abdomen6089.7893.627264.0784.67
9Pelvis5694.6897.705562.5383.12
10Pelvis6681.4695.143667.5582.08
Average gamma passing rate88.2295.2565.6380.76

Table 6 The MAE calculated from the predicted MLC positions and log files for the IMRT and VMAT cases

CaseBankPatient ID

12345678910
IMRTA (mm)1.04731.48931.84611.20761.99722.09382.12292.39521.35171.4592
B (mm)0.95291.19721.43051.24701.59241.56301.85162.15911.19171.5364
VMATA (mm)2.51631.95652.27884.29742.56802.75753.85892.04602.73763.0903
B (mm)2.74331.90082.60424.36622.40392.32584.04001.55702.78692.9853

References

  1. Brahme A. Optimization of stationary and moving beam radiation therapy techniques. Radiother Oncol. 1988;12:129-140.
    Pubmed CrossRef
  2. Otto K. Volumetric modulated arc therapy: IMRT in a single gantry arc. Med Phys. 2008;35:310-317.
    Pubmed CrossRef
  3. Ling CC, Burman C, Chui CS, Kutcher GJ, Leibel SA, LoSasso T, et al. Conformal radiation treatment of prostate cancer using inversely-planned intensity-modulated photon beams produced with dynamic multileaf collimation. Int J Radiat Oncol Biol Phys. 1996;35:721-730.
    Pubmed CrossRef
  4. Bortfeld T, Boyer AL, Schlegel W, Kahler DL, Waldron TJ. Realization and verification of three-dimensional conformal radiotherapy with modulated fields. Int J Radiat Oncol Biol Phys. 1994;30:899-908.
    Pubmed CrossRef
  5. Chui CS, LoSasso T, Spirou S. Dose calculation for photon beams with intensity modulation generated by dynamic jaw or multileaf collimations. Med Phys. 1994;21:1237-1244.
    Pubmed CrossRef
  6. Van Esch A, Bohsung J, Sorvari P, Tenhunen M, Paiusco M, Iori M, et al. Acceptance tests and quality control (QC) procedures for the clinical implementation of intensity modulated radiotherapy (IMRT) using inverse planning and the sliding window technique: experience from five radiotherapy departments. Radiother Oncol. 2002;65:53-70.
    Pubmed CrossRef
  7. Low DA, Moran JM, Dempsey JF, Dong L, Oldham M. Dosimetry tools and techniques for IMRT. Med Phys. 2011;38:1313-1338.
    Pubmed CrossRef
  8. Han Y. Review on the pre-treatment quality assurance for intensity modulated radiation therapy. Prog Med Phys. 2013;24:213-219.
    CrossRef
  9. Ezzell GA, Burmeister JW, Dogan N, LoSasso TJ, Mechalakos JG, Mihailidis D, et al. IMRT commissioning: multiple institution planning and dosimetry comparisons, a report from AAPM Task Group 119. Med Phys. 2009;36:5359-5373.
    Pubmed CrossRef
  10. Klein EE, Hanley J, Bayouth J, Yin FF, Simon W, Dresser S, et al; Task Group 142, American Association of Physicists in Medicine. Task Group 142 report: quality assurance of medical accelerators. Med Phys. 2009;36:4197-4212.
    Pubmed CrossRef
  11. Miften M, Olch A, Mihailidis D, Moran J, Pawlicki T, Molineu A, et al. Tolerance limits and methodologies for IMRT measurement-based verification QA: recommendations of AAPM Task Group No. 218. Med Phys. 2018;45:e53-e83.
    KoreaMed CrossRef
  12. Mehrens H, Taylor P, Followill DS, Kry SF. Survey results of 3D-CRT and IMRT quality assurance practice. J Appl Clin Med Phys. 2020;21:70-76.
    Pubmed KoreaMed CrossRef
  13. Ju SG, Han Y, Kum O, Cheong KH, Shin EH, Shin JS, et al. Comparison of film dosimetry techniques used for quality assurance of intensity modulated radiation therapy. Med Phys. 2010;37:2925-2933.
    Pubmed CrossRef
  14. Sharma M, Singh R, Dutt S, Tomar P, Trivedi G, Robert N. Effect of absorbed dose on post-irradiation coloration and interpretation of polymerization reaction in the Gafchromic EBT3 film. Radiation Physics Chem. 2021;187:109569.
    CrossRef
  15. McCurdy BM, Luchka K, Pistorius S. Dosimetric investigation and portal dose image prediction using an amorphous silicon electronic portal imaging device. Med Phys. 2001;28:911-924.
    Pubmed CrossRef
  16. Van Esch A, Depuydt T, Huyskens DP. The use of an aSi-based EPID for routine absolute dosimetric pre-treatment verification of dynamic IMRT fields. Radiother Oncol. 2004;71:223-234.
    Pubmed CrossRef
  17. Lee C, Menk F, Cadman P, Greer PB. A simple approach to using an amorphous silicon EPID to verify IMRT planar dose maps. Med Phys. 2009;36:984-992.
    Pubmed CrossRef
  18. Warkentin B, Steciw S, Rathee S, Fallone BG. Dosimetric IMRT verification with a flat-panel EPID. Med Phys. 2003;30:3143-3155.
    Pubmed CrossRef
  19. McCurdy BM, Greer PB. Dosimetric properties of an amorphous-silicon EPID used in continuous acquisition mode for application to dynamic and arc IMRT. Med Phys. 2009;36:3028-3039.
    Pubmed CrossRef
  20. Nelms BE, Rasmussen KH, Tome WA. Evaluation of a fast method of EPID-based dosimetry for intensity-modulated radiation therapy. J Appl Clin Med Phys. 2010;11:3185.
    Pubmed KoreaMed CrossRef
  21. Ansbacher W. Three-dimensional portal image-based dose reconstruction in a virtual phantom for rapid evaluation of IMRT plans. Med Phys. 2006;33:3369-3382.
    Pubmed CrossRef
  22. Rowshanfarzad P, Sabet M, O'Connor DJ, Greer PB. Improvement of Varian a-Si EPID dosimetry measurements using a lead-shielded support-arm. Med Dosim. 2012;37:145-151.
    Pubmed CrossRef
  23. Ko L, Kim JO, Siebers JV. Investigation of the optimal backscatter for an aSi electronic portal imaging device. Phys Med Biol. 2004;49:1723-1738.
    Pubmed CrossRef
  24. Rowshanfarzad P, Sabet M, O'Connor DJ, Greer PB. Reduction of the effect of non-uniform backscatter from an E-type support arm of a Varian a-Si EPID used for dosimetry. Phys Med Biol. 2010;55:6617-6632.
    Pubmed CrossRef
  25. Boylan C, McWilliam A, Johnstone E, Rowbottom C. The impact of continuously-variable dose rate VMAT on beam stability, MLC positioning, and overall plan dosimetry. J Appl Clin Med Phys. 2012;13:4023.
    Pubmed KoreaMed CrossRef
  26. Bertelsen A, Lorenzen EL, Brink C. Validation of a new control system for Elekta accelerators facilitating continuously variable dose rate. Med Phys. 2011;38:4802-4810.
    Pubmed CrossRef
  27. Nicolini G, Clivio A, Cozzi L, Fogliata A, Vanetti E. On the impact of dose rate variation upon RapidArc® implementation of volumetric modulated arc therapy. Med Phys. 2011;38:264-271.
    Pubmed CrossRef
  28. Baily NA, Horn RA, Kampp TD. Fluoroscopic visualization of megavoltage therapeutic x ray beams. Int J Radiat Oncol Biol Phys. 1980;6:935-939.
    Pubmed CrossRef
  29. Boon SN, van Luijk P, Schippers JM, Meertens H, Denis JM, Vynckier S, et al. Fast 2D phantom dosimetry for scanning proton beams. Med Phys. 1998;25:464-475.
    Pubmed CrossRef
  30. Ding L, Wu Q, Wang Q, Li Y, Perks RM, Zhao L. Advances on inorganic scintillator-based optic fiber dosimeters. EJNMMI Phys. 2020;7:60.
    Pubmed KoreaMed CrossRef
  31. Goiffon V, Virmontois C, Magnan P, Cervantes P, Corbière F, Estribeau M, et al. Radiation damages in CMOS image sensors: testing and hardening challenges brought by deep sub-micrometer CIS processes. Paper presented at: Proc. SPIE 7826, Sensors, Systems, and Next-Generation Satellites XIV, 78261S. Toulouse, France. 2010.
    CrossRef
  32. Tan J, Büttgen B, Theuwissen AJP. Radiation effects on CMOS image sensors due to X-Rays. Paper presented at: The Eighth International Conference on Advanced Semiconductor Devices and Microsystems. Smolenice, Slovakia: The Eighth International Conference on Advanced Semiconductor Devices and Microsystems. 2010: 279-282.
    KoreaMed CrossRef
  33. Huber A, Sergienko G, Kinna D, Huber V, Milocco A, Mercadier L. Response of the imaging cameras to hard radiation during JET operation. Fusion Eng Des. 2017;123:669-673.
    CrossRef
  34. Boon SN, van Luijk P, Böhringer T, Coray A, Lomax A, Pedroni E, et al. Performance of a fluorescent screen and CCD camera as a two-dimensional dosimetry system for dynamic treatment techniques. Med Phys. 2000;27:2198-2208.
    Pubmed CrossRef
  35. Petric MP, Robar JL, Clark BG. Development and characterization of a tissue equivalent plastic scintillator based dosimetry system. Med Phys. 2006;33:96-105.
    Pubmed CrossRef
  36. Collomb-Patton V, Boher P, Leroux T, Fontbonne JM, Vela A, Batalla A. The DOSIMAP, a high spatial resolution tissue equivalent 2D dosimeter for LINAC QA and IMRT verification. Med Phys. 2009;36:317-328.
    Pubmed CrossRef
  37. Guillot M, Beaulieu L, Archambault L, Beddar S, Gingras L. A new water-equivalent 2D plastic scintillation detectors array for the dosimetry of megavoltage energy photon beams in radiation therapy. Med Phys. 2011;38:6763-6774.
    Pubmed CrossRef
  38. Zin HM, Harris EJ, Osmond JP, Allinson NM, Evans PM. Towards real-time VMAT verification using a prototype, high-speed CMOS active pixel sensor. Phys Med Biol. 2013;58:3359-3375.
    Pubmed CrossRef
  39. Miao X, Sim T. Ambient image recovery and rendering from flash photographs. Paper presented at: IEEE International Conference on Image Processing 2005. Genova, Italy. 2005: II-1038.
    CrossRef
  40. Field R. .
  41. Wueller D. Evaluating digital cameras. Paper presented at: Proc. SPIE 6069, Digital Photography II, 60690K. San Jose, CA, USA. 2006.
    CrossRef
  42. Hsu WF, Chuang KW, Hsu YC. Comparisons of the camera OECF, the ISO speed, and the SFR of digital still-picture cameras. Paper presented at: Proc. SPIE 4080, Input/Output and Imaging Technolgies II. Taipei, Taiwan. 2000.
    CrossRef
  43. Hu Y, Zhu TC. Backscatter correction factor for megavoltage photon beam. Med Phys. 2011;38:5563-5568.
    Pubmed CrossRef
  44. Zhu TC, Ahnesjö A, Lam KL, Li XA, Ma CM, Palta JR, et al; AAPM Therapy Physics Committee Task Group 74. Report of AAPM Therapy Physics Committee Task Group 74: in-air output ratio, Sc, for megavoltage photon beams. Med Phys. 2009;36:5261-5291.
    Pubmed CrossRef
  45. Hesamian MH, Jia W, He X, Kennedy P. Deep learning techniques for medical image segmentation: achievements and challenges. J Digit Imaging. 2019;32:582-596.
    Pubmed KoreaMed CrossRef
  46. Zhou T, Ruan S, Canu S. A review: deep learning for medical image segmentation using multi-modality fusion. Array. 2019;3-4:100004.
    CrossRef
  47. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. Paper presented at: Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. Munich, Germany. 2015.
    CrossRef
  48. Ye JC, Han Y, Cha E. Deep convolutional framelets: a general deep learning framework for inverse problems. SIAM J Imaging Sci. 2018;11:991-1048.
    CrossRef
  49. Han Y, Ye JC. Framing U-Net via deep convolutional framelets: application to sparse-view CT. IEEE Trans Med Imaging. 2018;37:1418-1429.
    Pubmed CrossRef
  50. Ge Y, Wei D, Xue Z, Wang Q, Zhou X, Zhan Y, et al. Unpaired Mr to CT Synthesis with explicit structural constrained adversarial learning. Paper presented at: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). Venice, Italy. 2019: 1096-1099.
    Pubmed CrossRef
  51. Ge Y, Xue Z, Cao T, Liao S. Unpaired whole-body MR to CT synthesis with correlation coefficient constrained adversarial learning. Paper presented at: Proc. SPIE 10949, Medical Imaging 2019: Image Processing, 1094905. San Diego, CA, USA. 2019.
    CrossRef
  52. Lee D, Jeong SW, Kim SJ, Cho H, Park W, Han Y. Improvement of megavoltage computed tomography image quality for adaptive helical tomotherapy using cycleGAN-based image synthesis with small datasets. Med Phys. 2021;48:5593-5610.
    Pubmed CrossRef
  53. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. arXiv. 2015. https://doi.org/10.48550/arXiv.1512.03385.
    Pubmed CrossRef
  54. Isola P, Zhu JY, Zhou T, Efros AA. Image-to-image translation with conditional adversarial networks. Paper presented at: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI, USA. 2017.
    CrossRef
  55. Lasio G, Guerrero M, Goetz W, Lima F, Baulch JE. Effect of varying dose-per-pulse and average dose rate in X-ray beam irradiation on cultured cell survival. Radiat Environ Biophys. 2014;53:671-676.
    Pubmed CrossRef
  56. Posar JA, Davis J, Brace O, Sellin P, Griffith MJ, Dhez O, et al. Characterization of a plastic dosimeter based on organic semiconductor photodiodes and scintillator. Phys Imaging Radiat Oncol. 2020;14:48-52.
    Pubmed KoreaMed CrossRef
  57. Nyquist H. Certain factors affecting telegraph speed. Bell Syst Tech J. 1924;3:324-346.
    CrossRef
  58. Beck A, Teboulle M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J Imaging Sci. 2009;2:183-202.
    CrossRef
  59. Martinet B. [Brève communication. Régularisation d'inéquations variationnelles par approximations successives]. Revue française d'informatique et de recherche opérationnelle. Série rouge. 1970;4:154-158. French.
    CrossRef
  60. Combettes PL, Wajs VR. Signal recovery by proximal forward-backward splitting. Multiscale Model Simul. 2005;4:1168-1200.
    CrossRef
  61. Siddon RL. Fast calculation of the exact radiological path for a three-dimensional CT array. Med Phys. 1985;12:252-255.
    Pubmed CrossRef
  62. Heil CE, Walnut DF. Continuous and discrete wavelet transforms. SIAM Rev. 1989;31:628-666.
    CrossRef
  63. Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv. 2017. https://doi.org/10.48550/arXiv.1412.6980.
  64. Cheon W, Jung H, Lee M, Lee J, Kim SJ, Cho S, et al. Development of a time-resolved mirrorless scintillation detector. PLoS One. 2021;16:e0246742.
    Pubmed KoreaMed CrossRef
  65. Levitan E, Herman GT. A maximum a posteriori probability expectation maximization algorithm for image reconstruction in emission tomography. IEEE Trans Med Imaging. 1987;6:185-192.
    Pubmed CrossRef
  66. Biguri A, Dosanjh M, Hancock S, Soleimani M. TIGRE: a MATLAB-GPU toolbox for CBCT image reconstruction. Biomed Phys Eng Express. 2016;2:055010.
    CrossRef
  67. Low DA, Harms WB, Mutic S, Purdy JA. A technique for the quantitative evaluation of dose distributions. Med Phys. 1998;25:656-661.
    Pubmed CrossRef
  68. Depuydt T, Van Esch A, Huyskens DP. A quantitative evaluation of IMRT dose distributions: refinement and clinical assessment of the gamma evaluation. Radiother Oncol. 2002;62:309-319.
    Pubmed CrossRef
  69. Losasso T. IMRT delivery performance with a varian multileaf collimator. Int J Radiat Oncol Biol Phys. 2008;71:S85-S88.
    Pubmed CrossRef
  70. Li JG, Dempsey JF, Ding L, Liu C, Palta JR. Validation of dynamic MLC-controller log files using a two-dimensional diode array. Med Phys. 2003;30:799-805.
    Pubmed CrossRef
  71. Zygmanski P, Kung JH, Jiang SB, Chin L. Dependence of fluence errors in dynamic IMRT on leaf-positional errors varying with time and leaf number. Med Phys. 2003;30:2736-2749.
    Pubmed CrossRef
  72. Zeidan OA, Li JG, Ranade M, Stell AM, Dempsey JF. Verification of step-and-shoot IMRT delivery using a fast video-based electronic portal imaging device. Med Phys. 2004;31:463-476.
    Pubmed CrossRef
  73. Zeng GL. Maximum-likelihood expectation-maximization algorithm versus windowed filtered backprojection algorithm: a case study. J Nucl Med Technol. 2018;46:129-132.
    Pubmed CrossRef
  74. Hughes J. Automated analysis of Varian log files for advanced radiotherapy treatment verification: a multicenter study [dissertation]. Crawley: University of Western Australia; 2015.
  75. Manikandan A, Sarkar B, Holla R, Vivek TR, Sujatha N. Quality assurance of dynamic parameters in volumetric modulated arc therapy. Br J Radiol. 2012;85:1002-1010.
    Pubmed KoreaMed CrossRef
  76. Rangaraj D, Zhu M, Yang D, Palaniswaamy G, Yaddanapudi S, Wooten OH, et al. Catching errors with patient-specific pretreatment machine log file analysis. Pract Radiat Oncol. 2013;3:80-90.
    Pubmed CrossRef
  77. Agnew A, Agnew CE, Grattan MW, Hounsell AR, McGarry CK. Monitoring daily MLC positional errors using trajectory log files and EPID measurements for IMRT and VMAT deliveries. Phys Med Biol. 2014;59:N49-N63.
    Pubmed CrossRef
  78. Calvo-Ortega JF, Teke T, Moragues S, Pozo M, Casals-Farran J. A Varian DynaLog file-based procedure for patient dose-volume histogram-based IMRT QA. J Appl Clin Med Phys. 2014;15:4665.
    Pubmed KoreaMed CrossRef
  79. Bedford JL, Warrington AP. Commissioning of volumetric modulated arc therapy (VMAT). Int J Radiat Oncol Biol Phys. 2009;73:537-545.
    Pubmed CrossRef
Korean Society of Medical Physics

Vol.35 No.4
December 2024

pISSN 2508-4445
eISSN 2508-4453
Formerly ISSN 1226-5829

Frequency: Quarterly

Current Issue   |   Archives

Stats or Metrics

Share this article on :

  • line