Objective:To establish an automatic segmentation network based on different receptive fields for gross target volume (GTV) and organs at risk in patients with nasopharyngeal carcinoma.Methods:Radiotherapy data of 100 cases of nasopharyngeal carcinoma including CT images and GTV and organs at risk delineated by the physicians were collected. Ninety plans were randomly selected as the training dataset, and the other 10 plans as the validation dataset. Firstly, the images were subject to three data augmentation methods including center cropping, vertical flipping and rotation (-30°to 30°), and then input into MA_net networks proposed in this study for training. The model performance of networks was assessed by the number of network parameters (NP), floating-point number (FPN), the running memory (RM) and Dice index (DI), and eventually compared with DeeplabV3+ , PSP_net, UNet+ + and U_Net networks.Results:When the input image was in the size of 240×240, MA_net had a NP of 23.20%, 20.10%, 25.55% and 27.11% of these 4 networks, 50.02%, 19.86%, 6.37% and 13.44% for the FPN, 40.63%, 23.60%, 11.58% and 14.99% for the RM, respectively. For the DI of GTV, MA_net was 1.16%, 2.28%, 1.27% and 3.59% higher than these 4 networks. For the average DI of GTV and OAR, MA_net was 0.16%, 1.37%, 0.30% and 0.97% higher than these 4 networks.Conclusion:Compared with those four networks, the proposed MA_net network has slightly higher Dice index with fewer parameters, lower FPN and smaller RM.

Objective:To investigate the effects of CT images reconstructed using different field of view (FOV) sizes on the automatic segmentation of organs at risk and dose calculation accuracy in radiotherapy after radical mastectomy.Methods:Under the same scanning conditions, CT values-electron density conversion curves were established by reconstructing the original CT images of a phantom placed at the isocenter and extended FOV (eFOV) positions using FOV sizes of 50, 60, 70 and 80 cm. Then, these curves were compared. A standard phantom with a known volume was scanned, and the automatic segmentation result of the phantom on CT images reconstructed using different FOV sizes was compared. A total of 30 patients in Guangdong Second Provincial General Hospital from January 2020 to June 2022 with breast cancer were randomly selected. Through simulated positioning, their CT images were reconstructed using different FOV sizes for the purpose of automatic segmentation of organs at risk, followed by comparison between the outcomes of automatic segmentation and physicians′segmentation. The treatment plan established based on CT images reconstructed using a FOV size of 50 cm (FOV 50 images for short) was applied to CT images reconstructed using FOV sizes of 60, 70 and 80 cm (FOV 60, FOV 70 and FOV 80 images for short) for dose calculation, and the dose calculation result were compared. Results:The CT values - electron density conversion curves derived from CT images reconstructed using different FOV sizes were roughly consistent. At the isocenter, the difference between the segmented volume and actual volume of the standard phantom increased up to a maximum of 6 cm 3 (4.8%) with an increase in the FOV size. As indicated by the automatic segmentation result, the segmentation accuracy of the spinal cord, trachea, esophagus, thyroid, healthy mammary gland, and skin decreased with an increase in the FOV size ( t = -28.43-8.23, P < 0.05). The comparison of dose calculated based on CT images reconstructed using different FOV sizes showed that there was no statistically significant differences( P>0.05) in the dose to target volume ( V95) and the maximum and average doses in the supraclavicular lymph node region, as well as the dose to organs at risk. The coverage for planned target volume decreased with an increase in the FOV size, with a maximum difference of 4.06%. Conclusions:It is recommended that, for radiotherapy after radical mastectomy, FOV 50 images should be selected for the automatic segmentation of organs at risk, CT-values-electron density conversion curves should be established based on the electron density phantom images of the eFOV region, and the eFOV 80 images should be preferred for dose calculation.

Objective To propose a treatment planning optimization algorithm which can make full use of OAR dose distribution prediction meanwhile improving the output planning quality as much as possible.Methods We had reformulated an FMO function under the guidance of dose distribution prediction and also integrated equivalent uniform dose (gEUD) based on the consideration of prediction uncertainty,for providing optimal solution.Performance of the method was evaluated by comparing the optimized IMRT plan quality of 8 cervical cancers in the term of DVH curves,dose distribution and dosimetric endpoints with the original ones.Results The proposed method had a feasible,fast solution.Compared with original plan,its output plan had better plan quality in better dose homogeneity,less hot spot and further dose sparing for OARs.V30,V45 of rectum was decreased by (6.60±3.53)％ and (17.03±7.44)％,respectively,with the statistically significant difference (t=-4.954,-6.055,P＜0.05).V30,V45 of bladder was decreased by (14.74 ± 5.61) ％ and (14.99 ± 4.53) ％,respectively,with the statistically significant difference (t=-6.945,-8.759,P＜0.05).Conclusions We have successfully developed a predicted dose distribution and equivalent uniform dose-based planning optimization method,which is able to make good use of 3D dose prediction and ensure the output plan quality for intensity modulated radiation therapy.

Objective To establish a three-dimensional (3D) dose prediction model,which can predict multiple organs simultaneously in a single model and automatically learn the effect of the geometric anatomical structure on dose distribution.Methods Clinical radiotherapy plans of patients diagnosed with the same type of tumors were collected and retrospectively analyzed.For every plan,each organs at risk (OAR) voxel was regarded as the study sample and its deposited dose was considered as the dosimetric feature.A regularized multi-task learning method than could learn the relationship among different tasks was employed to establish the relationship matrix among tasks and the correlation between geometric structure and dose distribution among organs.In this experiment,the spinal cord,brainstem and bilateral parotids involved in the intensity-modulated radiotherapy (IMRT) plan of 15 nasopharyngeal cancer patients were utilized to establish the multi-organ prediction model.The relative percentage error between the predicted dose of voxel and the clinical planning dose was calculated to assess the feasibility of the model.Results Ten cases receiving IMRT plans were utilized as the training data,and the remaining five cases were used as the test data.The test results demonstrated a higher prediction accuracy and less data demand.And the average voxel dose errors among the spinal cord,brainstem and the left and right parotids were (2.01±0.02)％,(2.65± 0.02) ％,(2.45± 0.02) ％ and (2.55± 0.02) ％,respectively.Conclusion The proposed model can accurately predict the dose of multiple organs in a single model and avoid the establishment of multiple single-organ prediction models,laying a solid foundation for patient-specific plan quality control and knowledge-based treatment planning.

Objective:To establish a correlation model between MRI and CT images to generate synthetic-CT (sCT) of head and neck cancer during MRI-guided radiotherapy by using generative adversarial networks (GAN).Methods:Images and IMRT plans of 45 patients with nasopharyngeal carcinoma were collected before treatment. Firstly, the MRI (T1) and CT images were preprocessed, including rigid registration, clipping, background removal and data enhancement, etc. Secondly, the cases were trained by GAN, of which 30 cases were randomly selected and put into the network as training set images for modeling and learning, and the other 15 cases were used for testing. The image quality of predicted sCT and real CT were statistically compared, and the dose distribution recalculated upon predicted sCT was statistically compared with that of real planned dose distribution.Results:The mean absolute error of the predicted sCT of the testing set was (79.15±11.37) HU, and the SSIM value was 0.83±0.03. The MAE values of dose distribution difference at different regional levels were less than 1% compared to the prescription dose. The gamma passing rate of the sCT dose distribution was higher than 92% and 98% under the 2mm/2% and 3mm/3% criteria.Conclusions:We have successfully proposed and realized the generation of sCT for head and neck cancer using GAN, which lays a foundation for the implementation of MRI-guided radiotherapy. The comparison of image quality and dosimetry shows the feasibility and accuracy of this method.

Objective:To compare the accuracy and generalized robustness of three predictive models of knowledge-based treatment strategies for radiotherapy for optimized model selection.Methods:The clinical radiotherapy plans of 45 prostate cancer (PC) cases and 25 nasopharyngeal cancer (NPC) cases were collected, and analyzed using three models (Z, L and S model), proposed by Zhu et al, Appenzoller et al and Shiraishi et al, respectively, to predict the dose-volume histogram (DVH) of bladder and rectum on PC cases and that of left and right parotid on NPC cases. The prediction error was measured by the difference of area under the predicted DVH and the clinical DVH curves (|V (pre_DVH)-V (clin_DVH)|), where a smaller prediction error implies a greater prediction accuracy. The accuracies of these three models were compared on the single organ at risk (OAR), and the generalized robustness of models was evaluated and compared by calculating the standard deviation of the prediction accuracy on different OAR. Results:For bladder and rectum, the prediction error of L model (0.114 and 0.163, respectively) was significantly higher than those values of Z and S models (≤0.071, P<0.05); for left parotid gland, the predicted error of S model (0.033) did not present significant difference from those values of Z and L models (≤0.025, P>0.05); for right parotid gland, S model (0.033) demonstrated significantly higher prediction error than those of Z and L models (≤0.028, P<0.05). Regarding different OAR, S model showed a lower standard deviation of prediction accuracy when comparing to Z and L models (0.016, 0.018 and 0.060, respectively). Conclusions:In the prediction of DVH in bladder and rectum of PC, Z and S models were more accurate than L model. In contrast, Z and L models demonstrated higher accuracy than S model in the prediction of left and right parotid glands of NPC. In respect to different OAR, the generalized robustness of S model was superior than the other two models.