Deep-learning-based Detection and Segmentation of Organs at Risk in Head and Neck

Abstract

At present, one of the most widely used cancer treatment methods is radiotherapy. A key step in developing a radiotherapy plan is to accurately delineate all at-risk organs (OARs) to minimize adverse effects on surrounding healthy organs. However, manual delineating by experienced doctors is limited by the degree of experience and fatigue, time-consuming and error prone. The speed and accuracy of traditional sketching methods are very low, which can not meet the clinical application very well. In this paper, a method based on deep learning is proposed for the detection and segmentation of head and neck dangerous organs. 48 groups of data in PDDCA public dataset are used to train and test the network. The design idea is divided into two networks (1) OAR detection network (2) OAR segmentation network, The training method is a three-step strategy, (1) feature extraction, (2) OAR detection network, and (3) ROI region segmentation. Adopt the idea of less training data and more testing data. In the end, under the training of small data, the segmentation results close to the world’s top level were still obtained, and after testing with multiple sets of data, the generalization ability and learning potential of the network were verified.