Research On Deep Learning-based Dose Distribution Prediction Of Radiotherapy Plan For Head And Neck Cancer

Head and neck tumors are the sixth most common malignant tumors worldwide,accounting for about 20%-30%of all malignant tumors.There are approximately 500,000 new cases of head and neck cancer worldwide each year,and the incidence is showing a clear upward trend,especially among female populations.At present,radiotherapy has become the main treatment option in clinical practice.The key goal of radiotherapy is to kill cancerous cells and protect surrounding normal tissues by ensuing that the prescribed dose is delivered to the planning target volume(PTV)while minimizing the dose deposited in surrounding normal organs-at-risk(OARs)as much as possible.In the plan design of radiotherapy for head and neck cancer,determining a good dose distribution is the key issues,and it is also one of the important factors of radiotherapy effect.However,traditional radiotherapy plan design is difficult to ensure the consistency of plan quality and to provide individualized plans.Therefore,it is of great clinical significance to study the automatic design of dose distribution of radiotherapy plan in practical radiotherapy.The research goal of this paper is to achieve automatic and accurate prediction of three-dimensional dose prediction in head and neck cancer radiotherapy plan.The research content is to utilize the deep learning techniques to design an efficient and accurate dose regression model.Specifically,utilizing the publicly available OpenKBP-2020 Challenge datasets of head and neck caner radiotherapy plans.We propose a novel field-wise dose decomposition learning method to predict dose distribution map from global to local to global,which consists of two stages.The first stage is global dose learning,which predicts rough global dose distribution maps using CT images through Global Dose Network(GDN);and the second stage is local dose refinement,which introduces field segmentation masks as prior information,calculates loss function using the dose-volume parameters of regions-of-interest(ROIs)and 3D gradient map extracted by Sobel operator,to further refine the dose distribution prediction in local regions.The experimental results mainly include two parts:1)Quantitative results:Compared with the top-ranked method in the OpenKBP-2020 Challenge,our proposed method achieves the highest prediction accuracy,with a 31.6%improvement in Dose score and a 35.4%improvement in DVH score.Furthermore,the predicted accuracy of our propose method in the direction of the radiation field is significantly higher than that of the current state-of-the-art(SOTA)methods.In addition,the difference between the predicted dose-volume histogram(DVH)dosimetry parameters obtained by our proposed method and the real DVH dosimetry parameters is the smallest;2)Visualized results:Compared with other SOTA methods,the prediction of three-dimensional dose distribution map by our proposed method is closer to the corresponding ground truth,especially in local areas of fields,ROIs and dose boundary.Furthermore,the DVH curves predicted by our proposed method are closer to the real DVH curves,demonstrating that the model has better predictive performance in areas of ROIs.In conclusion,the study provides a valuable reference for the automatic design of radiotherapy plans for head and neck cancer,and is expected to be applied in actual clinical radiotherapy.