2490 - An Adaptive Multi-Branch UNet Model for Intelligent Dose Prediction and Treatment Planning Optimization in Cervical Cancer Radiotherapy

Purpose/Objective(s): To address the challenge of balancing multiple loss functions in the multi-branch UNet (MB-UNet), an adaptive multi-branch UNet model is proposed.

Materials/Methods: Based on the intelligent radiotherapy planning system KylinRay-TPSe independently developed by SuperAccuracy Science & Technology Co. Ltd., it is proposed an Adaptive Multi-Branch UNet (AMB-UNet) dose prediction model for cervical cancer. Firstly, input features are convolved at different scales to distinguish dense features from sparse features, and channel attention is used to integrate outputs at each encoder layer of the model to achieve more effective feature extraction. Secondly, a reasonable loss function selection and adaptive weight combination strategy are used during training to improve the robustness of the model. Thirdly, the target region is expanded, and overlapping areas are removed from the corresponding OAR contours to improve prediction in critical areas. Finally, once the dose distribution is predicted, the predicted dose is used as an optimization objective to construct an objective function for optimization, omitting constraint settings and making it easier to find a local minimum that meets the optimization conditions. This study selected a cervical cancer dataset of 220 cases with prescription doses between 45 Gy and 49.95 Gy, including 160 cases for training, 30 for validation, and 30 for testing.

Results: On the test set, the mean absolute error (MAE) of the dose predicted by the AMB-UNet model was 1.18 Gy, with a relative prescription error of 2.5% (compared to 1.35 Gy and 2.86% for the MB-UNet model). The average absolute errors for D1, D5, and D95 of the target region were 0.59 Gy, with a relative prescription error of 1.2%. For some key dose-volume metrics, the predicted values were close to the actual values: the predicted relative volume for bladder V45 was (27.47 ± 5.93)%, and the actual was (27.57 ± 6.09)%; for rectum V45, the predicted was (35.3 ± 9.0)% and the actual was (35.86 ± 8.05)%; for small intestine V40, the predicted was (5.48 ± 4.02)% and the actual was (5.47 ± 3.72)%. In terms of plan optimization, the optimized dose met clinical constraint requirements. Compared with the original plan, while ensuring target dose coverage, the dose to some OARs was also reduced. Especially considering the rapid dose falloff outside the target area in actual plans, the dose outside the target could be further reduced by additional automatic ring generation around the target and setting penalty terms in the objective function. For example, the relative volume of bladder V45 was reduced by an average of 1.23%.

Conclusion: The AMB-UNet method can achieve fast and accurate dose prediction in cervical cancer radiotherapy and has good application prospects in clinical radiotherapy planning, significantly improving the planning efficiency for physicists.