Research On Application Of Deep Learning Method In Dose Prediction Distribution Of Radiation Therapy

The number of patients with malignant tumors in China is increasing year by year,and malignant tumors have become a major disease threatening human life and health.Tumor radiotherapy has become one of the important methods for clinical treatment of tumors due to its high cure rate and wide range of suitable diseases.Before the implementation of radiotherapy,clinical tumor radiotherapy doctors and physicists need to design and optimize tumor radiotherapy plans to ensure that the cancer cells are killed to the greatest extent while the radiation dose to normal tissues is minimized.Knowledge-Based Planning(KBP)is a data-driven radiotherapy planning method that uses a large number of previously optimized plans to predict new plans,which has been a research hotspot in recent years.With the development of deep learning,the method of using deep learning to predict the dose distribution of tumor radiotherapy has demonstrated its superiority.Different from traditional methods,the deep learning-based dose prediction method can quickly and accurately predict the radiation dose distribution of tumor patients without manually extracting features.This thesis is based on the Open KBP challenge dataset held by the 62 nd American Association of Physicists in Medicine(AAPM)in 2020,and explores the application of deep learning methods in three-dimensional dose prediction.The main research results are as follows:(1)Dose prediction neural network structure improvement method: an improved U-Net network for radiotherapy three-dimensional dose distribution is proposed.The influence of convolutional neural network structure in deep learning on the prediction performance is explored,including residual module,attention mechanism module and cascade structure,and the results show the effectiveness of these improvements.The average absolute error of the single U-Net model is 2.619 Gy.The cascade U-Net network combined with the residual module and the convolutional block attention module(CBAM)can reduce the average absolute dose prediction error to 2.442 Gy while increasing the amount of parameters by approximately 4.5 times.(2)Dose prediction neural network architecture search method: Aiming at the limitation of manually designing the convolutional neural network structure,a gradient-based neural network architecture search framework named U-NAS is proposed,which can automatically and efficiently search for the structure with the best performance.The average absolute error of the U-NAS single model is 2.597 Gy,and the error of its cascade network is 2.407 Gy.(3)Dose prediction neural network adversarial learning-based ensemble method:Aiming at the problem of high learning time and computational cost using traditional ensemble,an ensemble framework adopting knowledge distillation based on adversarial learning named KDA-Net is proposed,which uses the searched U-NAS models as the teacher models,and U-Net as the student model.The parameter amount and prediction time of the student model do not change with the number of teacher models used for ensemble.As a result,KDA-Net reduces the average absolute dose prediction error of the single U-Net model from 2.619 Gy to 2.565 Gy.