Objective:To construct a general deep learning dose prediction model applicable to radiotherapy for head and neck tumors, establish design methods for artificial intelligence (AI)-assisted radiotherapy plan and evaluate the accuracy of prediction.Methods:Radiotherapy plans of 818 patients who received radiotherapy for head and neck cancers from January 2018 to June 2021 in Cancer Hospital of Chinese Academy of Medical Sciences were enrolled. Patients involved 17 types of common head and neck cancers, and the prescribed dose covered 5 kinds of dose gradients ranging from 54 Gy to 73.92 Gy. And 1-2 cases per each cancer type (31 cases in total) were randomly selected as the validation set, and the remaining 787 cases were used as the training set to build a deep learning head and neck radiotherapy generalized dose prediction model. Then based on the dose prediction results of this model, a program was written to automatically generate inverse optimization condition scripts, which were sent back to the treatment planning system to achieve AI-assisted radiotherapy plan design. Among the patients who received radiotherapy in our hospital from June 2021 to January 2022, 1 patient for each disease type (17 cases in total) was selected to evaluate the AI-assisted plan design program and evaluate its clinical feasibility using paired t-test. Results:Dose prediction model accuracy evaluation revealed that in the 31-case validation set, there was no statistical difference in the evaluation metrics of clinical concern for organs at risks, except for the D 1 cm3 prediction for spinal cord planning risk volume, which was statistically different compared with the clinical reference plan. The AI-assisted plan design program had higher plan quality metric scores (37.88±6.42) than manual plans (35.00±7.63) in 17 test cases ( t=-1.00, P=0.166). The number of manual adjustments to the inverse optimization conditions was reduced from (5.47±2.97) times to (2.76±1.00) times for the AI-assisted plan compared to the manual-only plan ( t=4.12, P<0.001). And the number of outlined dose shaping structures was reduced from 7.35±3.98 to 3.12±1.18 ( t=5.61, P<0.001). Conclusions:The unified universal model of dose prediction established for different head and neck cancers has high accuracy in dose prediction for all types of head and neck tumor plans. The AI-assisted planning method established in this pattern can reduce the clinical workload of physicists and improve the efficiency of their work.

Objective:To assess intrafractional errors of the deformation of clinical target volumes (CTVs) during online adaptive radiotherapy (OART) for prostate cancer patients, aiming to provide a basis for online plan optimization.Methods:A retrospective analysis was conducted for 13 prostate cancer patients who received 1.5 T magnetic resonance imaging (MRI)-guided OART (8 Gy × 5 fractions, totaling 65 fractions) at the Department of Radiation Oncology, Cancer Hospital, Chinese Academy of Medical Sciences. The MRI images were collected at the beginning and end of various treatment fractions. Then, the CTVs and organs at risk (OARs) were delineated by the same radiation oncologist. After rigid registration and triangle mesh generation, the surface vertices were extracted. The deformable registration for the CTV surfaces was performed using the thin-plate spline robust point matching (TPS-RPM) algorithm, yielding vertex correspondences. Last, both systematic and random intrafractional errors of CTV deformation were calculated.Results:The average Hausdorff distance (HD) for deformable registration of treatment fractions was (1.68 ± 0.28) mm. The intrafractional systematic errors of CTV deformation were (0.25 ± 3.18) mm (anterior-posterior direction; A-P), (0.89 ± 3.85) mm (left-right direction, L-R), and (-1.98 ± 6.69) mm (superior-inferior direction, S-I). The intrafractional random errors of CTV deformation were determined at (-0.26 ± 1.89) mm (A-P), (-0.08 ± 0.88) mm (L-R), and (-0.04 ± 1.86) mm (S-I).Conclusions:During OART, CTV deformations primarily occur in the S-I direction. Therefore, it is necessary to consider the expanded size of margins in this direction during OART for prostate cancer.

Objective:To assess intrafractional errors of the deformation of clinical target volumes (CTVs) during online adaptive radiotherapy (OART) for prostate cancer patients, aiming to provide a basis for online plan optimization.Methods:A retrospective analysis was conducted for 13 prostate cancer patients who received 1.5 T magnetic resonance imaging (MRI)-guided OART (8 Gy × 5 fractions, totaling 65 fractions) at the Department of Radiation Oncology, Cancer Hospital, Chinese Academy of Medical Sciences. The MRI images were collected at the beginning and end of various treatment fractions. Then, the CTVs and organs at risk (OARs) were delineated by the same radiation oncologist. After rigid registration and triangle mesh generation, the surface vertices were extracted. The deformable registration for the CTV surfaces was performed using the thin-plate spline robust point matching (TPS-RPM) algorithm, yielding vertex correspondences. Last, both systematic and random intrafractional errors of CTV deformation were calculated.Results:The average Hausdorff distance (HD) for deformable registration of treatment fractions was (1.68 ± 0.28) mm. The intrafractional systematic errors of CTV deformation were (0.25 ± 3.18) mm (anterior-posterior direction; A-P), (0.89 ± 3.85) mm (left-right direction, L-R), and (-1.98 ± 6.69) mm (superior-inferior direction, S-I). The intrafractional random errors of CTV deformation were determined at (-0.26 ± 1.89) mm (A-P), (-0.08 ± 0.88) mm (L-R), and (-0.04 ± 1.86) mm (S-I).Conclusions:During OART, CTV deformations primarily occur in the S-I direction. Therefore, it is necessary to consider the expanded size of margins in this direction during OART for prostate cancer.

Artificial intelligence (AI) technologies are being widely applied in radiotherapy. However, the integration of AI into clinical workflows of radiotherapy faces a series of challenges, such as poor model interpretability, domain shifts between clinical application and training data, and the inherent model uncertainties. Therefore, case-specific quality assurance (QA) is essential before deploying AI models in clinical practice. This paper reviews and summarizes QA methodologies for the application of AI models in radiotherapy across four key areas: image registration, image generation, region of interest segmentation, and treatment planning.

Objective:To construct a general deep learning dose prediction model applicable to radiotherapy for head and neck tumors, establish design methods for artificial intelligence (AI)-assisted radiotherapy plan and evaluate the accuracy of prediction.Methods:Radiotherapy plans of 818 patients who received radiotherapy for head and neck cancers from January 2018 to June 2021 in Cancer Hospital of Chinese Academy of Medical Sciences were enrolled. Patients involved 17 types of common head and neck cancers, and the prescribed dose covered 5 kinds of dose gradients ranging from 54 Gy to 73.92 Gy. And 1-2 cases per each cancer type (31 cases in total) were randomly selected as the validation set, and the remaining 787 cases were used as the training set to build a deep learning head and neck radiotherapy generalized dose prediction model. Then based on the dose prediction results of this model, a program was written to automatically generate inverse optimization condition scripts, which were sent back to the treatment planning system to achieve AI-assisted radiotherapy plan design. Among the patients who received radiotherapy in our hospital from June 2021 to January 2022, 1 patient for each disease type (17 cases in total) was selected to evaluate the AI-assisted plan design program and evaluate its clinical feasibility using paired t-test. Results:Dose prediction model accuracy evaluation revealed that in the 31-case validation set, there was no statistical difference in the evaluation metrics of clinical concern for organs at risks, except for the D 1 cm3 prediction for spinal cord planning risk volume, which was statistically different compared with the clinical reference plan. The AI-assisted plan design program had higher plan quality metric scores (37.88±6.42) than manual plans (35.00±7.63) in 17 test cases ( t=-1.00, P=0.166). The number of manual adjustments to the inverse optimization conditions was reduced from (5.47±2.97) times to (2.76±1.00) times for the AI-assisted plan compared to the manual-only plan ( t=4.12, P<0.001). And the number of outlined dose shaping structures was reduced from 7.35±3.98 to 3.12±1.18 ( t=5.61, P<0.001). Conclusions:The unified universal model of dose prediction established for different head and neck cancers has high accuracy in dose prediction for all types of head and neck tumor plans. The AI-assisted planning method established in this pattern can reduce the clinical workload of physicists and improve the efficiency of their work.

Artificial intelligence (AI) technologies are being widely applied in radiotherapy. However, the integration of AI into clinical workflows of radiotherapy faces a series of challenges, such as poor model interpretability, domain shifts between clinical application and training data, and the inherent model uncertainties. Therefore, case-specific quality assurance (QA) is essential before deploying AI models in clinical practice. This paper reviews and summarizes QA methodologies for the application of AI models in radiotherapy across four key areas: image registration, image generation, region of interest segmentation, and treatment planning.

Objective To improve the performance of auto-segmentation of prostate target area and organs-at-risk in online magnetic resonance image and enhance the efficiency of magnetic resonance imaging-guided adaptive radiotherapy(MRIgART)for prostate cancer.Methods A retrospective study was conducted on 40 patients who underwent MRIgART for prostate cancer,including 25 in the training set,5 in the validation set,and 10 in the test set.The planning CT images and corresponding contours,along with online MR images,were registered and input into a deep learning network for online MR image auto-segmentation.The proposed method was compared with deformable image registration(DIR)method and single-MR-input deep learning(SIDL)method.Results The overall accuracy of the proposed method for auto-segmentation was superior to those of DIR and SIDL methods,with average Dice similarity coefficients of 0.896 for clinical target volume,0.941 for bladder,0.840 for rectum,0.943 for left femoral head and 0.940 for right femoral head,respectively.Conclusion The proposed method can effectively improve the accuracy and efficiency of auto-segmentation in MRIgART for prostate cancer.

Objective To evaluate the effects of the multiple factors especially image geometric accuracy of the imaging system on the segmentations of target areas and organs-at-risk.Methods The study used phantoms to test the imaging performance of the 1.5T magnetic resonance(MR)linear accelerator system,including the assessments of MR image geometric distortion and the segmentation errors caused by factors such as image geometric distortion.Model 604-GS large field MR image distortion phantom was used to explore the geometric distortion of the MR images for MR-guided radiotherapy;and CIRS Model 008z upper abdominal phantom was used to analyze the segmentation errors of target areas and organs-at-risk.Results The average geometric distortion and maximum distortion of 3D T1WI-FFE images vs 3D T2WI-TSE images were 0.54 mm vs 0.53 mm and 1.96 mm vs 1.68 mm,respectively;and the control points of the large distortions were distributed at the edges of the phantom,which was consistent with the MR imaging characteristics previously reported.Compared with CT-based segmentation contour,the MDA was 1.17 mm and DSC was 0.91 for 3D T1WI-FFE,while MDA was 0.86 mm and DSC was 0.94 for 3D T2WI-TSE.Conclusion The study quantitatively assesses the geometric accuracy of the imaging system for MR-guided radiotherapy.The phantom-based contour analysis reveals that with CT image as gold standard,the segmentation error in MRI images meets the clinical requirements,and that 3D T2WI-TSE image is advantageous over 3D T1WI-FFE image in segmentation accuracy.

Using high-energy proton to image the region of interest can directly obtain the accurate estimation of the proton stopping power of the lesions,which is of great significance to reduce the range uncertainty in proton therapy.As a fundamental function of proton computed tomography(CT),radiographic imaging plays a crucial role in assisting clinical positioning.The study develops a compact proton CT detector based on an active array pixel CMOS chip in Monte-Carlo simulation toolkit Geant4,and evaluates the radiographic imaging capability of the system using 180 MeV protons.The angles of tracks are successfully reconstructed.CTP404,CTP528,and the CTP515 of specific materials are used for simulation,obtaining the spatial and density resolutions,and measuring the proton relative stopping power(RSP).The image signal-to-noise ratio is improved when using 2° proton scattering angle cut-off value.The spatial resolution is 3-4 lp/cm measured using CTP528 module.The density resolution is better than 0.05 g/cm3,and the RSP resolution is within 5%when CTP404 module is used.Through the imaging of CTP515 phantom of specific material,it is demonstrated that the system has potential for imaging common human tissues.

Objective:To evaluate the feasibility of using MRI-only simulation images for dose calculation of both photon and proton radiotherapy for nasopharyngeal carcinoma cases.Methods:T 1-weighted MRI images and CT images of 100 patients with nasopharyngeal carcinoma treated with radiotherapy in Cancer Hospital of Chinese Academy of Medical Sciences from January 2020 to December 2021 were retrospectively analyzed. MRI images were converted to generate pseudo-CT images by using deep learning network models. The training set, validation set and test set included 70 cases, 10 cases and 20 cases, respectively. Convolutional neural network (CNN) and cycle-consistent generative adversarial neural network (CycleGAN) were exploited. Quantitative assessment of image quality was conducted by using mean absolute error (MAE) and structural similarity (SSIM), etc. Dose assessment was performed by using 3D-gamma pass rate and dose-volume histogram (DVH). The quality of pseudo-CT images generated was statistically analyzed by Wilcoxon signed-rank test. Results:The MAE of the CNN and CycleGAN was (91.99±19.98) HU and (108.30±20.54) HU, and the SSIM was 0.97±0.01 and 0.96±0.01, respectively. In terms of dosimetry, the accuracy of pseudo-CT for photon dose calculation was higher than that of the proton plan. For CNN, the gamma pass rate (3 mm/3%) of the photon radiotherapy plan was 99.90%±0.13%. For CycleGAN, the value was 99.87%±0.34%. The gamma pass rates of proton radiotherapy plans were 98.65%±0.64% (CNN, 3 mm/3%) and 97.69%±0.86% (CycleGAN, 3 mm/3%). For DVH, the dose calculation accuracy in the photon plan of pseudo-CT was better than that of the proton plan.Conclusions:The deep learning-based model generated accurate pseudo-CT images from MR images. Most dosimetric differences were within clinically acceptable criteria for photon and proton radiotherapy, demonstrating the feasibility of an MRI-only workflow for radiotherapy of nasopharyngeal cancer. However, compared with the raw CT images, the error of the CT value in the nasal cavity of the pseudo-CT images was relatively large and special attention should be paid during clinical application.