Deep Learning Methods For Automatic Target Delineation And Dose Prediction Of Radiotherapy

Objective: Cancer is one of the main causes of human death,which seriously hinders the improvement of human life expectancy.Radiotherapy is a common method of cancer treatment.The process of making and implementing radiotherapy plan manually is heavy and time-consuming,and depends on subjective clinical experience.The rise of artificial intelligence technology has brought more possibilities to solve the above problems.The radiotherapy planning and implementation process based on artificial intelligence model has the advantages of higher efficiency and better consistency.Therefore,this topic will explore the new artificial intelligence technology applied to radiotherapy planning.Materials and methods: The main application scenarios of artificial intelligence technology in radiotherapy include automatic segmentation,automatic planning,generation of computed tomography（CT）images and automatic quality assurance（QA）.Among them,the clinical demand for automatic segmentation and automatic planning is greater.Therefore,this article will use deep learning technology to realize the automatic segmentation of target area and the dose prediction of the radiotherapy plan before the implementation of radiotherapy,so as to improve the efficiency and accuracy of radiotherapy.This subject takes nasopharyngeal carcinoma as an example.For the target delineation task,we designed a multimodal information fusion module based on the attention mechanism,which uses the information of MRI to assist the tumor segmentation in CT images.192 patients with nasopharyngeal carcinoma were included in this experiment.The data of these patients included CT images,MRI and tumor masks.The dataset was divided into training,validation and test dataset,including 134,20 and 38 patients respectively.The performance of domain alignment module and multi-modal information fusion segmentation network are verified respectively.For the dose prediction task,we designed the distance transformation module,and introduced the distance prior information into the model to improve the accuracy of dose prediction.This study included 201 patients with nasopharyngeal carcinoma,40 of whom were from another hospital and will be used as an external test dataset.The dataset of these patients include plan CT,masks of regions of interest and clinical plan.The internal dataset was divided into training,validation and test dataset,including 130,11 and 20 patients respectively.The dose prediction performance with and without distance transformation module is compared.Results: For the target delineation task,it was determined that T2 images could assist the tumor segmentation in CT images through the pre-experiment.Then,the tumor segmentation was performed in CT and T2 images.The Dice similarity coefficient（DSC）,95% Hausdorff distance（HD95）and average symmetry surface distance（ASSD）of the tumor segmentation results based on the registered multimodal images are 0.7181,9.6637 mm and 2.8014 mm respectively.The DSC,HD95 and ASSD of the tumor segmentation results based on the generated multimodal images are 0.7062,9.7620 mm and 2.8882 mm respectively,which were better than the tumor segmentation results of CT images（0.7020,10.0757 mm and 3.0472mm）on the whole.For the dose prediction task,we evaluate the predicted dose distribution and the inverse planning results respectively.The dose error and dose volume histogram（DVH）error of the dose distribution predicted by our proposed model are 7.51% and 11.6% lower than the mask-based model,respectively.The performance of the inverse plan obtained by our proposed model is similar to that of the mask-based model on nasopharynx gross tumor volume（GTVnx）and organs-at-risk,and is better on lymph node gross tumor volume（GTVnd）and clinical target volume（CTV）.The passing rate is increased from 89.490%and 90.016% to 96.694% and 91.189% respectively.Conclusion: The above target delineation and dose prediction models have been verified by experiments and show great performance,which brings new research directions for artificial intelligence assisted radiotherapy planning and new possibilities for automatic and efficient radiotherapy planning and implementation.