Ca-Mt: A Self-Ensembling Model For Semi-Supervised Cardiac Segmentation With Elliptical Descriptor Based Contour-Aware

Nowadays, semi-supervised learning has achieved greatly in computer aided cardiac segmentation, which utilized unlabeled data to cut annotation cost and improve the network performance. However, most existing methods neglect the contour constrain of cardiac which often lead to inaccurate segmentation result, particularly around boundaries. In this paper, we proposed a novel self-ensembling approach for cardiac segmentation to leverage unlabeled data and contour information of labeled data. To achieve this, we develop an elliptical contour descriptor to describe the segmentation contour in several sets of parameters. Furthermore, a contour-aware loss is designed to minimize the difference between target and prediction. To leverage the annotated data, we then introduce a self-ensembling method named teacher-student network which requires the prediction of teacher and student branches being consistent under random perturbation or transformation of input images. We integrate our contour-aware method into this framework and let it gradually learn the contour information from coarse to fine. Experiments on the 2017 MICCAI Automated Cardiac Diagnosis Challenge (ACDC) and Left Atrial Segmentation Challenge 2013 (LASC'13) datasets show that our approach improves the cardiac segmentation performance and outperforms existing semi-supervised segmentation approaches.