Self-supervised regularization for abdominal multi-organ segmentation

Zhengshan Huang

Keywords: Semi-supervised Learning, Abdominal Multi-organ Segmentation, Efficient Context-Aware Network

TL;DR: Self-supervised regularization for abdominal multi-organ segmentation

Abstract: Automated segmentation of abdominal multi-organ from 3D computed tomography images (CTs) is necessary for organ quantification, surgical planning, and disease diagnosis. Manual delineation practices require anatomical knowledge, are expensive, time consuming and can be inaccurate due to human error. Here, we describe a semi-supervised efficient context-aware segmentation network for abdominal multi-organ segmentation from 3D CTs based on encoder-decoder architecture. For learning more useful feature information from unlabeled cases, a variational auto-encoder branch is added to reconstruct the input image itself in order to regularize the shared encoder and impose additional constraints on its layers. And for the purpose of consuming less source, an efficient context-aware segmentation backbone network is used in this paper.