A Novel Deep Learning Framework for Standardizing the Label of OARs in CT

Abstract

When organs at risk (OARs) are contoured in computed tomography (CT) images for radiotherapy treatment planning, the labels are often inconsistent, which severely hampers the collection and curation of clinical data for research purpose. Currently, data cleaning is mainly done manually, which is time-consuming. The existing methods for automatically relabeling OARs remain unpractical with real patient data, due to the inconsistent delineation and similar small-volume OARs. This paper proposes an improved data augmentation technique according to the characteristics of clinical data. Besides, a novel 3D non-local convolutional neural network is proposed, which includes a decision making network with voting strategy. The resulting model can automatically identify OARs and solve the problems in existing methods, achieving the accurate OAR re-labeling goal. We used partial data from a public head-and-neck dataset (HN_PETCT) for training, and then tested the model on datasets from three different medical institutions. We have obtained the state-of-the-art results for identifying 28 OARs in the head-and-neck region, and also our model is capable of handling multi-center datasets indicating strong generalization ability. Compared to the baseline, the final result of our model achieved a significant improvement in the average true positive rate (TPR) on the three test datasets (+8.27%, +2.39%, +5.53%, respectively). More importantly, the F1 score of small-volume OAR with only 9 training samples increased from 28.63% to 91.17%.