Research And Application Of Deep Learning-Based Methods For Organs-at-Risk Auto-Delineation And Dose Prediction In Radiation Therapy

Radiotherapy is one of the most important means of cancer treatment today.However,radiotherapy for tumors is facing major challenges due to the high difficulty and complexity of the process.The entire process of tumor radiotherapy,from initial diagnosis and treatment plan development to treatment plan implementation,heavily relies on the expertise of doctors.However,the level of treatment in hospitals varies greatly,and there is high heterogeneity,which,along with the increasing incidence of cancer,leads to a shortage of excellent tumor doctors and physicists in small and medium-sized hospitals,making it difficult to improve the treatment effect.The rise of artificial intelligence technology has provided a better solution and development direction for addressing the uneven distribution of medical resources and developing accurate radiotherapy plans.Therefore,this thesis takes the application of artificial intelligence technology in the radiotherapy process as the main focus and carries out research in the following three parts to address the current difficulties and pain points in the research field:(1)We propose an end-to-end segmentation model to address the automatic organ delineation problem in radiotherapy planning by fusing edge and geometric information.Firstly,the model combines a Transformer with a 3D convolutional neural network,which effectively utilizes the advantages of modeling local contextual information by convolutional neural networks and learning global semantic relevance by Transformer.Secondly,to address the problem of edge blurring caused by low image contrast,an edge-aware and edge-guided feature module is designed to enhance the extraction of organ contour information during training.In addition,to alleviate the segmentation performance bottleneck caused by large differences in organ shape between patients,a multi-task training paradigm is introduced to the proposed model,which adds an auxiliary organ shape regression task to improve the model’s ability to capture shape information.Finally,to address the problem of class imbalance between large and small organs,a label distribution calibration strategy is designed to encourage the model to focus more on small target organ learning during training and improve the final segmentation performance of the model.(2)We propose a multi-stage 3D radiotherapy dose distribution prediction framework based on difficulty-aware learning.To address the challenging of voxel-wise regression task,we introduce a novel multi-stage end-to-end framework,which decomposes the task into more manageable sub-tasks for progressively predicting the radiation dose.In addition,a difficultyaware learning mechanism is proposed to continuously focus on hard-to-predict regions during the training process,using a voxel-level attention-weighted mean absolute error loss.Furthermore,to avoid the time-consuming hyperparameter tuning process for the loss function,we propose a loss adaptive learning strategy to automatically balance the contributions of different stages of the model to the final dose prediction results,achieving more accurate dose prediction performance.(3)We develop a tumor radiotherapy planning assistance tool in the process of algorithm research.It enables oncologists,physicists,and researchers to quickly implement services such as organ contouring and dose prediction during radiotherapy in a concise and intuitive interactive form.The software integrates five core functions,including medical image visualization and contouring result review,autonomous training for tasks such as critical organ segmentation and radiotherapy dose prediction,an AI-based question and answer assistant,as well as remote server access and transmission tools.Ultimately,the software and deep learning algorithms achieve precision radiotherapy,enhancing the efficiency and information level of the radiotherapy department.In summary,the automatic segmentation and dose prediction models proposed in this thesis are experimentally validated on a publicly available head and neck dataset.The experimental results show that the proposed methods perform well compared to existing methods,indicating their potential to further promote research on AI-assisted radiation therapy planning and support the precision and intelligence development of cancer radiation therapy.