Beyond General Edge Utilization: Edge Attention Mean Teacher for Semi-Supervised Medical Image Segmentation

Abstract

Deep learning has achieved substantial success in the field of semi-supervised medical image segmentation. Current researches mainly concentrate on enhancing pseudo-label generation process and refining consistency regularization architectures. However, the edge information, which is essential for medical image segmentation but scarce in semi-supervised scenario, is often overlooked. To address this problem, we present an edge attention mean teacher (EAMT) method that goes beyond general edge extraction to better leverage the edge information for improved segmentation performance. Particularly, based on a novel definition of edge, we propose a new edge extraction method to boost the edge extraction capability of model. Furthermore, we elaborately design an edge-aware loss function that uses the extracted edges as additional supervision for labeled data and as masks for unlabeled data. The EAMT method is characterized by its capability to extract and leverage robust edge information to promote the learning process for both labeled and unlabeled data. We evaluate the segmentation performance of the proposed EAMT method on two public 3D datasets (LA and Pancreas-CT). Experimental results demonstrate that EAMT achieves superior segmentation performance compared to several state-of-the-art methods in semi-supervised medical image segmentation.