CycleGan-based kV-to-MV image translation for potential in vivo dosimetry application: concept proposal and cross-institutional validation

Abstract

Background

This work aims to develop and validate a novel CycleGan-based methodology to transfer the kV planning CT (pCT) to the reference MV portal images, potentially applicable to in vivo treatment dose monitoring.

Methods

The kV projections of pCT were prepared based on the various gantry angles of MV projections using treatment beams on Varian Halcyon system. A CycleGAN-based network incorporating attention module (ECA-CycleGAN) was trained to learn the relationship between kV and MV images, which performance was compared with the conventional Pix2pix and CycleGAN methods quantitatively. The beam angles and multi-leaf collimator parameters retrieved from the clinical plans were used to segment the treatment apertures on the model-generated reference MV images, within which the sensitivity to the artificial errors were tested using gamma analysis. Cross-institutional validations were performed on multiple machines and scanning protocols.

Results

Comparing the 2574 model-generated MV images with the measured ground truth of 13 validation cases, the mean ± standard deviation of the structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR) and root mean square error (RMSE) were 0.969 ± 0.007, 41.3 ± 3.2 and 1.6·10^− 2±3.8·10^− 4 for ECA-CycleGAN, consistently better than that of using CycleGAN (0.940 ± 0.004, 35.7 ± 3.3 and 2.6·10^− 2±4.5·10^− 4) and pix2pix (0.931 ± 0.005, 32.4 ± 4.0 and 4.2·10^− 2±4.8·10^− 4) respectively. After introducing artificial translational or rotational errors, the gamma passing rates decreased and the gamma indices increased significantly (all P < 0.05). The ECA-CycleGAN model displayed good generalizability across various pCT scanners, imaging protocols and Halcyon accelerators from 2 institutions.

Conclusion

Without complex and time-consuming Monte Carlo simulations, the proposed ECA-CycleGAN network facilitates the efficient establishment of the reference MV portal images applicable to in vivo transmitted dosimetry. It may potentially improve the accuracy of dose delivery especially for the advanced treatment techniques when pretreatment measurement verification or inter-fractional dose remediation are impossible.

Background

Intensity modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) can deliver a more conformal dose to tumor target while sparing the surrounding normal tissues [1]. Considering the sharp dose fall-off and technical complexity of IMRT/VMAT, patient-specific plan quality assurance (QA) is usually performed prior to the treatment delivery to prevent possible errors [2]. As a typical procedure of plan QA, the measured dose distribution in a phantom is compared with planned results in treatment planning system (TPS) to ensure the consistency of the two [3]. However, pre-treatment phantom-based dosimetric verification ignores the dose deviation induced by both machine errors and patient changes during the treatment, such as inter-fractional and intra-fractional anatomic variations, hence cannot reflect the actual patient dose [4]. Moreover, measurement-based pre-treatment QA is not possible for on-line adaptive radiotherapy (ART) when patient is immobilized on the treatment table during the whole process [5, 6]. Using the images of the treatment day, the delivered dose on the patient anatomies can be retrospectively reconstructed, supporting dose-guided ART to correct the previous errors in the upcoming fractions [7, 8]. However, compared with the retrospective approaches, real-time in vivo dose monitoring is clinically desirable to prevent errors more promptly, especially for less-fractionated techniques without opportunities for remediation, such as stereotactic radio-surgery (SRS) [9].

High-spatial resolution, good linear dose response and fast image acquisition speed are the advantages of electronic portal imaging device (EPID) [10–13], facilitating exit-beam based real-time in vivo dose monitoring during the treatment delivery. However, as an essential premise, it is technically difficult and time-consuming to develop a reference expected EPID response to verify the measured exit beam. Previous studies have predicted the reference image by convolving the fluence map with a dose kernel, which can be either simulated by Monte Carlo (MC) method [14] or calculated by a series of empirical data [15]. The clinical application of these methods are impeded by the additional calibration work [4] and potential influence of different dose response between water and the EPID material [16]. MC method can accurately predict the reference images by comprehensively simulating the particle transport and interactions in the accelerator, patient, EPID panel and supporting device [17]. However, MC calculation is time-consuming and requires cutting-edge hardware and software, which is clinically not affordable when the patient volume is large. Moreover, the unavailable industrial secrets such as detailed geometric and material data of linear accelerator (LINAC) and EPID structures may prevent the accurate implementation of those simulations.

Deep learning (DL) has been demonstrated as an accurate and efficient method for predicting medical images in other applications [18–21], which is potentially applicable to in vivo dosimetry. Zhang et al. [22] developed a combined DL and MC model to simplify the process of simulating the EPID response. Bosco et al. [23] developed an alternative computational approach for EPID-based non-transit dosimetry using a U-net model. Jia et al. [24] reconstruct a 3D dose distribution as imparted to a virtual cylindrical phantom using an electronic portal imaging device (EPID) for pre-treatment patient-specific quality assurance.

Cycle-Consistent Generative Adversarial Network (CycleGAN) [25] enabled image-to-image translation between two domains without requiring paired datasets. Although it has been widely used to generate CT images from cone beam CT (CBCT) and magnetic resonance imaging (MRI) data [26, 27], studies on synthesizing MV projections from planning CT remain scarce especially for the Halcyon/Ethos (Varian Medical Systems, Palo Alto, CA) systems. Compare with the aforementioned studies using conventional C-arm LINAC and simulated dose image as the input of the DL model, Halcyon/Ethos systems are potentially more suitable and necessary for in vivo dosimetry due to: (1) The EPID is fixed on the beam stopper rather than retractable robot arms, minimizing the geometric uncertainties induced by gravities of various gantry angles; (2) The EPID is sealed in the gantry bore without potential collision hazard during the treatment delivery, which also records the exit beam of the full treatment size in 28 cm* 28 cm; (3) All Halcyon/Ethos systems are beam-matched globally sharing identical dosimetric models in the TPS, potentially facilitating the inter-institutional technical transferability and clinical applicability of the developed in vivo dosimetry method; (4) Halcyon utilizes identical 6 MV photon beams for both treatment and MV cone beam CT (MVCBCT) imaging [28], providing more direct and available data for model training than using MC simulations. However, MC simulation for Halcyon/Ethos is currently difficult if not possible, considering some necessary technical details have not been revealed publicly by the vendor, such as the material component of the new dual-layer multi-leaf collimator (MLC); (5) The online ART technique performed on Ethos prevents the conventional measurement-based pre-treatment QA, which highlights the value of in vivo dose monitoring during treatment delivery. However, no previous study has focused on the kV-MV image translation potentially application to in vivo dosimetry for the relatively new Halcyon/Ethos system by far.

Based on patient data acquired on multiple imaging and treatment machines from two institutions, this study developed and validated an improved novel CycleGAN network to predict the reference EPID responses to the exit treatment beams using the projections of planning kV CT at the corresponding gantry angles, which can be potentially applied to real-time in vivo dosimetry by comparing the instantly measured and pre-developed reference EPID images.

Materials and methods

Patient data

This study was approved by Institutional Review Board with informed consent signed by the patients (2021-04-151-K01). The inclusion criteria were (1) Pathologically confirmed primary, recurrent, or metastatic head and neck cancer; (2) Available planning CT (pCT) and treatment plan information; (3) Treated on the Halcyon platform using MVCBCT image guidance. The exclusion criteria were (1) Obvious anatomic changes or artificial implants after surgery; (2) Poor performance of thermoplastic mask and other immobilization methods; (3) pCT and MVCBCT images with severe artifacts and low resolution.

From two institutions (A and B), 33 patients were selected and immobilized using thermoplastic masks before pCT acquisition. The 20 patients in the training cohort were scanned using 120 kVp on pCT scanner A (Siemens Somatom Confidence Scanner, Siemens Healthineers, Erlangen, Germany). The field size, resolution and slice thickness of the pCT A were 50 cm × 50 cm, 512 pixels×512 pixels and 3 mm respectively. In total, 13 patients were used in the cross-protocol and cross-institutional generalizability test. Specifically, 9 validation patients were selected from institution A using the same pCT A protocols as the training cohort. Another 3 validation patients were selected using a different tube voltage (100 kVp) on the pCT scanner A. From institution B, 1 validation patient was selected using a different pCT scanner B (Siemens Somatom go.Top Scanner, Siemens Healthineers, Erlangen, Germany) and a different Halcyon machine. The field size, resolution and slice thickness of the pCT B were 70 cm × 70 cm, 512 pixels×512 pixels and 3 mm respectively. The detailed pCT scanning and treatment information for the 13 validation patients are shown in Table 1.

Table 1.

Patient characteristic in the validation cohort

Patient Index	pCT kVp	Treatment Technique	Number of Fields	Control Points in each field
A1	120	3DCRT	2	300
A2	120	IMRT	9	166
A3	120	IMRT	4	166
A4	120	IMRT	9	166
A5	100	IMRT	6	166
A6	120	IMRT	6	166
A7	120	IMRT	9	166
A8	120	IMRT	9	166
A9	120	IMRT	9	166
A10	120	IMRT	9	166
A11	100	VMAT	2	178
A12	100	VMAT	2	178
B13	120	VMAT	2	178

Note: pCT represents planning CT; 3DCRT represents three-dimensional conformal radiation therapy; IMRT represents intensity-modulated radiotherapy; VMAT represents volumetric-modulated arc therapy. Patients A1-A12 are from institution A and Patient B13 is from institution B

All the patients were treated on the beam-matched Halcyon platforms using identical 6 MV photon beams for both treatment and MVCBCT guidance. Therefore, the grayscale of region-of-interest (ROI) in the MVCBCT projection is proportional to that of the treatment beamlet [10–13]. The MLC positions segmenting the apertures at each gantry angle and various control points were retrieved in the treatment plans [29]. The original MVCBCT projections before reconstruction were used in this study.

Pre-processing of projection images

As shown in the workflow (Fig. 1), a rigid registration was conducted on pCT to match the field-of-view (FOV) of MVCBCT using ITK-SNAP software (version 4.0) [30]. The registered pCT images were cropped to fit the FOV of MVCBCT in image domain. Consistent with the geometries of MVCBCT projections, the pCT images were projected into various angles of cone-beam CT geometries using Tomographic Iterative GPU-based Reconstruction (TIGRE) [31]. In total, there were 4040 paired projections in the training cohort and 2574 paired projections in the test cohort.

Fig. 1.

The workflow of the study

ECA-CycleGAN

As an unpaired image-to-image translation method, this work improved the conventional CycleGAN [28] by adding the efficient channel attention (ECA) block in extracting features module. As a channel attention mechanism, ECA was used to balance the paradox of performance and complexity trade-off [32]. For each feature map, the aggregated features were gained using global average pooling (GAP). After 1D convolution using a kernel size adaptively determined by the channel dimension C, a Sigmoid function was used to learn the channel attention. The and b were constants.

To compare the performance of the improved ECA-CycleGAN network, another 2 DL models using pix2pix and conventional CycleGAN were parallelly trained to translate the pCT projection domain into MVCBCT projection domain [33]. To learn the mapping matrix between the two image domains, the dataloader of the original image-to-image translation methods loaded the image file formats (the suffix. ‘.jpg’, ‘.png’, ‘.bmp’, etc.) whose pixels were integers ranging from 0 to 255. However, the pixels of the log normalized projections were singles ranging from 0 to 10. Therefore, it was necessary to transfer them into image file formats before normalization and standardization of input module. In order to decrease the loss of converting raw files to image files, the dataloader of the three networks were modified to load the original projections (the suffix ‘.mat’) rather than the processed images.

The size of input CT and MVCBCT projection images were both 1280*1280 with a voxel width of 0.336 mm. They were initially resized to 256*256 before the training process. Each network was trained for 90 epochs using learning rate of 0.0002 and batch size of 2. The training and testing were performed on a NVIDA RTX 3070 with 16 GB of RAM.

Analysis of image consistency

Using the whole patient anatomies displayed in the measured MV projections as true values, the following evaluation metrics were used to compare the accuracy of the model-generated MV images using kV pCT. The performance of pix2pix, CycleGAN and ECA-CycleGAN methods were compared quantitatively.

As a measurement of disparities, root mean square error (RMSE) of each detected and generated image pair was defined as:

where M, N was the width and height of each projection. and was the value of different pixel.

As a measurement of consistencies, the peak signal-to-noise ratio (PSNR) was defined as:

where MAX was the maximum pixel value of the image.

As a measurement of similarity between the detected and generated MV projections, the structural similarity index measure (SSIM) [34] was defined as:

where x and y represented the detected projection and generated projection; and were the average of x and y; and were the variances of x and y, respectively; was the covariance of x and y; and were constants.

According to these definitions, better model performance can be suggested by the larger PSNR value, lower MSE value, and the SSIM value closer to 1 [35].

Test of sensitivity to artificial errors

To test the sensitivity of the proposed method to the artificial errors simulating patient deviations in the treatment fields, translational and rotational errors were introduced into the 3D volumes of planning CT respectively and generated the MV images with setup errors. The gamma indices and gamma passing rates were compared between the MV images before and after the geometric deviations were introduced, using the validation cases listed in the Table 1. Using the method proposed by Low et al. [36] and the recommendation of the American Association of Physicists in Medicine Task Group 307 report (AAPM TG-307) [37], the 2 mm/3% criteria [38] were adopted to calculate the gamma index and gamma passing rates in the ROI segmented on the measured and model-generated projections respectively, in accordance with the beamlet apertures defined by the MLC parameters extracted from the treatment plans.

To simulate the exit beam errors induced by translational deviations of patient anatomies, the isocenters were shifted along the transverse, sagittal and coronal axis by 2 mm simultaneously. To simulate the transmitted dose errors induced by patient rotation, the registered pCT images were rotated by 2 in transverse, sagittal and coronal planes, simultaneously. The errors were simulated by changing the value of ‘offOrigin’ in TIGRE and using Pillow (version 9.5.0) in Python (version 3.7.12). As an empirical validation of the proposed method, the head and neck region of an anthropomorphic phantom (CIRS, Computerized Imaging Reference Systems, USA) was also used. Limited by the non-rotatable couch of Halcyon/Ethos, the rotational errors including arbitrary shift (9 mm) and rotation (3°) errors were manually introduced to simulate the geometric deviations of the phantom.

The gamma indices and gamma passing rates of the validation patients were calculated as mean ± standard deviation. To compare the differences before and after the introduction of artificial errors, t test was performed on the gamma indices and gamma passing rates using Python (version 3.7.12). A P-value of two-sided hypothesis test smaller than 0.05 was considered as statistically significant.

Results

Network performance

It took 1 d 11 h 2 min, 1 d 14 h 6 min and 1 d 15 h 54 min to train the ECA-CycleGAN, CyclyGAN and Pix2pix models respectively. It took less than 1 s to predict a reference MV image from a kV pCT projection, using any one of the trained deep learning networks.

As a visual comparison, Fig. 2 shows the kV projections of pCT, the MVCBCT projections and model-generated MV projections of the same beam angle. The difference images were also calculated by subtracting the measured from the generated MV images. Each row shows the images of an example patient: patient A2 treated with IMRT (Fig. 2A-D), patient A8 treated with VMAT (Fig. 2E-H), and patient B13 treated with VMAT (Fig. 2I-L).

Fig. 2.

The projection images of 3 example patients: patient A2 treated with IMRT (A-D), patient A8 treated with VMAT (E-H), and patient B13 treated with VMAT (I-L). The difference images (D, H, L) were computed by subtracting the MVCBCT projections (B, F, J) from the generated MV images (C, G, K) using the kV projections of planning CT (pCT) at corresponding gantry angles

Based on the 13 validation patients as shown in Table 1, the quantitative consistency between the MVCBCT projections and the model-generated MV projections within the patient anatomies without the air background were compared in Table 2. The validation results were presented as mean ± standard deviation of 1814 image pairs from 9 patients using 120 kV pCT at institution A, 606 image pairs from 3 patients using 100 kV pCT at institution A and 154 image pairs from 1 patient from institution B, respectively. The performance of the proposed ECA-CycleGAN method was compared with the conventional Pix2pix and CycleGAN networks.

Table 2.

Mean and standard deviation of SSIM, PSNR and RMSE of different networks

	SSIM	PSNR(dB)	RMSE
Institution A (1814 projections from 9 patients using 120 kV pCT)
ECA-CycleGAN	0.969 ± 0.006	41.4 ± 3.1	1.7·10− 2±3.8·10− 4
CycleGAN	0.939 ± 0.004	35.2 ± 3.2	2.9·10− 2±4.3·10− 4
pix2pix	0.930 ± 0.005	32.0 ± 3.8	4.2·10− 2±4.5·10− 4
Institution A (606 projections from 3 patients using 100 kV pCT)
ECA-CycleGAN	0.974 ± 0.004	44.3 ± 2.9	1.2·10− 2±2.8·10− 4
CycleGAN	0.941 ± 0.004	37.4 ± 3.2	2.6·10− 2±4.1·10− 4
pix2pix	0.935 ± 0.002	33.6 ± 3.4	3.8·10− 1±6.5·10− 4
Institution B (154 projections from 1 patient using different pCT and Halcyon machines)
ECA-CycleGAN	0.954 ± 0.002	38.5 ± 2.7	1.8·10− 2±3.1·10− 4
CycleGAN	0.938 ± 0.001	34.3 ± 2.8	2.7·10− 2±3.9·10− 4
pix2pix	0.933 ± 0.002	32.2 ± 3.1	2.9·10− 2±7.5·10− 4
All (2574 projections from 13 patients using various pCT and Halcyon machines)
ECA-CycleGAN	0.969 ± 0.007	41.3 ± 3.2	1.6·10− 2±3.8·10− 4
CycleGAN	0.940 ± 0.004	35.7 ± 3.3	2.6·10− 2±4.5·10− 4
pix2pix	0.931 ± 0.005	32.4 ± 4.0	4.2·10− 2±4.8·10− 4

Note: SSIM represents the structural similarity index measure; PSNR represents the peak signal-to-noise ratio; RMSE represents the root mean square error

Sensitivity test to artificial errors

All 13 validation patients from institution A and B were used to test the sensitivity to various artificial errors. Table 3 shows the statistics of Gamma analyses in the treatment beam apertures segmented on the ECA-CycleGAN generated MV projections using the field angles and MLC apertures of the clinical plans. As empirical validations of the potential application to in vivo dosimetry, Table 3 also shows the Gamma analyses of the anthropomorphic CIRS phantom between the MV projections before and after simulated translational and rotational errors were introduced.

Table 3.

Statistics of gamma passing rates and gamma indices in the apertures of treatment beams using the ECA-CycleGAN generated MV projections. Translational or rotational errors (2 mm offset or 2° rotation on the transverse, sagittal and coronal planes) were simulated to test the sensitivity of the proposed method to the transmitted dose deviations induced by patient anatomies. The results were presented as mean ± standard deviation

	Institution A 120 kVp CT (5969 beamlets)	Institution A 100 kVp CT (694 beamlets)	Institution B (138 beamlets)	Phantom (without background)
Gamma passing rates
Reference	95.65%±4.72%	97.81%±4.09%	94.55%±6.18%	100%
Offset	87.81%±17.61% (P<0.001)	95.83%±6.81% (P<0.001)	92.25%±7.88% (P=0.014)	89.22%±4.64% (P<0.001)
Rotation	90.02%±11.43% (P<0.001)	95.04%±8.50% (P<0.001)	87.85%±11.66% (P<0.001)	69.30%±14.64% (P=0.006)
Gamma indices
Reference	0.37±0.29	0.37±0.30	0.24±0.36	0
Offset	0.41±0.33 (P<0.001)	0.40±0.33 (P<0.001)	0.26±0.35 (P<0.001)	0.40±0.08 (P=0.016)
Rotation	0.43±0.34 (P<0.001)	0.42±0.35 (P<0.001)	0.27±0.40 (P<0.001)	0.76±0.25 (P=0.022)

To simulate the in vivo dose monitoring during the treatment delivery, based on the same example patients of Figs. 2 and 3 displays the impact of the artificial errors on the gamma passing rates and gamma indices within the treatment apertures segmented on the generated MV images using the proposed ECA-CycleGAN network. Each subfigure represents a specific gantry angle, where the beam apertures were defined by the various control points of related MLCs. The first and second columns were affected by translational errors. The third and fourth columns were affected by rotational errors. The second and forth columns displayed the profiles of gamma indices along the inline (green lines) and crossline (blue lines). The horizontal red lines on the profiles indicate the gamma index equal to one, the voxels above which failed gamma analysis. The locations of the failed areas were marked as red in the first and third columns. The points with value of -1 were out of the MLC-shaped treatment field and were not considered.

Fig. 3.

The cases of in vivo dosimetry corresponding to different induced errors, based on the same example patient examples of Fig. 2. (a1-3) Generated MV images with the translational errors, with red points indicating not passing areas. (b1-3, c1-3) Crossline and inline gamma index profiles of generated MV images with the translational errors. (d1-3) Generated MV images with the rotational errors, with red points indicating not passing areas. (E, F) Crossline and inline gamma index profiles of generated MV images with the rotational errors

Discussion

DL has emerged as a promising method for medical image translation tasks. Baroudi et al. [39] employed CycleGAN and cGAN models to predict MVCBCT images and effectively enhanced the visualization of pacemakers. Scholey et al. [40] generated MVCT images from MRI using a fully convolutional 3D U-Net architecture, demonstrating the feasibility of MRI-MVCT only treatment planning. This work further demonstrated that kV-to-MV image translation method is potentially applicable to the in vivo dosimetry for precision radiotherapy.

The in vivo dosimetry provides additional confidence to the successful delivery of advanced treatment techniques such as online ART or SBRT, potentially preventing accidents more promptly instead of taking remedial actions afterwards based on retrospective dose reconstruction, hence received an increasing clinical and research interest. Jia et al. [24] reconstructed a 3D dose distribution by using a DL-based signal processing strategy, however, it was only validated on a virtual cylindrical phantom. Although Zhang et al. [22] also utilized DL network to predict the EPID portal image based on C-arm LINACs, their method relied on MC simulation for dose image prediction, which is less efficient and more complex. In addition, there has been no publication reporting transmitted dosimetry for Halcyon/ Ethos based on MC methods, probably due to publicly unavailable technical details from the vendor. In contrast, our network provided reference maps consistent with the planned dose without complex MC process, which is potentially applicable to in vivo dosimetry by comparing with the measured exit beams during the treatment delivery. Favored by the identical MV beams used for both treatment and MVCBCT imaging, the proposed DL-based transmitted dosimetry method avoided the long computational time, complex algorithm and calibration measurement compared with the conventional MC simulation [22] or convolution methods [41]. It took about 1 min in total to generate 202 kV projections from pCT and convert them to reference MV images using the proposed model, which is clinically affordable in terms of efficiency, hardware and technologic complexity compared with MC method, especially when the patient volume is large.

The machine-specific pCT kV energy spectrum and MV beam commission for the conventional C-arm LINACs also impede the clinical implementation of transmitted dosimetry, considering a model developed for one machine/institution is hard if not impossible to be applied to the other. In contrast, the globally beam-matched Halcyon systems does not only facilitate the data collection for model training but also favors the implementation of the developed models across institutions. Therefore, good generalizability was observed across various Halcyon LINACs, as shown in Table 2; Fig. 2. Moreover, the DL models also displayed good robustness amongst different pCT scanners and kV tube energies, which solved a major challenge of applying DL models in different institutions.

As illustrated in Table 2, all 3 DL networks successfully converted the kV pCT projections to EPID response to the MV exit beam. By incorporating the channel attention into DL networks, the ECA-CycleGAN network was proposed to solve the image to image translation problem using an unpaired training strategy [25], achieving better consistency with the measured ground truth than that of the conventional CycleGAN and pix2pix models, as suggested by all evaluation metrics including SSIM, PSNR and RMSE. The superior performance of ECA-CycleGAN network was also consistently observed across various pCT scanners, kV tube energies and Halcyon machines. As possible explanations, the network with the ECA module could better extract the significant features than the conventional CycleGAN model, especially in nasal areas [42]. Considering the variation of anatomical structures, set-up settings, and couches between pCT and MV projections increased the difficulty and uncertainties of registration, pix2pix based on paired images was less accurate in this task which focused more on the patient specific features yet lacking generalization ability.

As shown in Table 3; Fig. 3, the proposed method can alarm the artificial errors quantitatively in terms of significantly decreased gamma passing rates and increased gamma indices (all p < 0.05). The localization of failed areas can be displayed on the reference images, overlaying the aperture of the treatment beamlet. The profiles of gamma indices can be updated simultaneously with the customized placement of crosshairs in Fig. 3, providing a convenient tool for the clinicians to evaluate the error magnitude and location before making a decision. Again, the proposed method displayed good generalizability across different institutions, applicable on various pCT scanners, kV tube energies and Halcyon machines. Considering the 3D treatment couch of Halcyon/Ethos can only correct translational errors, the in vivo dose monitoring is clinically valuable to detect potentially unacceptable deviations induced by rotational errors. Moreover, the proposed method is also applicable to detect the dose deviations induced by inter- and intra-fractional anatomical deformations. In addition to patient anatomies, the discrepancies between the reference MV images and exit treatment beams are also partially contributed by the delivery parameters. Therefore, the proposed method is complementary to the log file-based QA approaches as reported by previous studies [43, 44] and commercial products such as MobiusFX (Mobius Medical Systems, Houston, TX) [45] and LINACWatch^® (Qualiformed, La Roche-sur-Yon, France) [46].

As for limitation of this study, the proposed method has not been tested on patients by comparing the detected in vivo transmitted dosimetry during the treatment, because the EPID mounted on commercial Halcyon systems records the integrated response instead of frame-by-frame cine mode exit beams. These experiments are potentially doable on the research machines with iTools license [47] or when this engineering problem is solved by the vendor in the future.

Conclusion

This work proposed a novel ECA-CycleGAN network to generate the reference EPID images potentially applicable to in vivo treatment dose monitoring, outperforming the conventional CycleGAN and pix2pix networks in accuracy, and prevailing over the Monte-Carlo method in efficiency and accessibility. As multi-institutional validation, the ECA-CycleGAN model-generated MV images agreed well with the measured true values using various pCT scanners, kV tube-voltage settings and different Halcyon LINACs. The feasibility and sensitivity of the proposed transmitted dosimetry method was tested by introducing artificial errors to clinical treatment plans, displaying good generalizability across institutions.

Abbreviations

pCT

Planning CT

IMRT

Intensity modulated radiation therapy

VMAT

Volumetric modulated arc therapy

QA

Quality assurance

TPS

Treatment planning system

ART

Adaptive radiotherapy

SRS

Stereotactic radio-surgery

EPID

Electronic portal imaging device

MC

Monte Carlo

LINAC

Linear accelerator

DL

Deep learning

MLC

Multi-leaf collimator

ROI

Region-of-interest

FOV

Field-of-view

TIGRE

Tomographic Iterative GPU-based Reconstruction

ECA

Efficient channel attention

GAP

Global average pooling

3DCRT

Three-dimensional conformal radiation therapy

RMSE

Root mean square error

PSNR

Peak signal-to-noise ratio

SSIM

Structural similarity index measure

Author contributions

YZ, XL and JQ are the chief investigators, and conceived and designed the study; RQ and ZS developed the deep learning model, analyzed results and wrote original draft; QLa and YZ collected data, analyzed results and revised the manuscript; HH, ZD, SY, JL, JD, GM, CY, QLc, FX, TL, CL collected and assembled the images data. (Superscript letters following author names indicate their institutional affiliations (e.g., QLa and QLc), as detailed in the Title page.)

Funding

This research was jointly funded by National Natural Science Foundation of China (12475309, 12275012, 12411530076, 12581360004, 82202941); Beijing Natural Science Foundation (Z210008); Fundamental Research Funds for the Central Universities/ Clinical Medicine Plus X - Young Scholars Project of Peking University (PKU2025PKULCXQ014); Ministry of Education Exchange Program for Teachers and Students of Higher Education Institutions in the Chinese Mainland/Hong Kong/Macao (7111400072); National Key R&D Program of China (2019YFF01014405).

Data availability

The raw data supporting the conclusions of this article will be made available upon email request.

Declarations

Ethics approval and consent to participate

This study was approved by Ethics Committee of Deyang People’s Hospital with informed consent signed by the patients (2021-04-151-K01) and adhered to the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.