Accurate Three Dimensional Radiotherapy Dose Prediction For Esophageal Cancer Based On Deep Learning

Treatment planning is one of the most important steps in esophageal radiotherapy,which aims at achieving the maximal dose delivery at the planning target volume(PTV)while reducing dosage at surrounding organ-at-risks(OARs).Due to the complexity in clinical practice,treatment planning often involves tedious manual adjustment till the radiotherapy plan is clinical accepted.Accurate dose prediction can reduce manual adjustments by providing close to optimal dose distributions,which can substantially improve both the planning efficiency and accuracy.Compared with other diseases,the location,shape and scales of PTV and OARs in esophageal cancer are highly inconsistent.Besides,the treatment parameters such as the number and direction of radiation beams could be different between each plan,which poses challenges to design robust dose prediction algorithms.To address the issues mentioned above,this paper studies on developing accurate three-dimensional dose prediction algorithm for esophageal radiotherapy based on deep learning,which includes: 1)Designing feature maps including a signed-distance image as model input,which highlights the geometric distribution and spatial correlation of each PTV and OARs;2)Constructing a high-resolution multi-scale convolutional neural network based on dilated convolution to preserve the details of small objects(such as spinal cord)and extract both global and local features from multiple scales;3)Improving the model performance on multi-center data by introducing three-dimensional cascade mechanism,embedded parameters of radiation field,data augmentations and advanced optimization strategies to training procedure.The proposed method is evaluated on esophageal data set and outperforms other state-of-the-art algorithm by a large margin.Besides,the method is also evaluated on other disease data(such as head and neck cancer),and the results indicate proposed model has great generalization ability,which could provide a powerful tool for the research of intelligent radiotherapy planning.