Contour-aware consistency for semi-supervised medical image segmentation

In medical images, the edges of organs are often blurred and unclear. Existing semi-supervised image segmentation methods rarely model edges explicitly. Thus most methods produce inaccurate predictions in target edge regions. In this paper, we propose a contour-aware consistency framework for semi-supervised medical image segmentation. The framework consists of a shared encoder, a vanilla primary decoder and a contour-enhanced auxiliary decoder. The contour-enhanced decoder is designed to enhance the features of the target contour region. The predictions from the primary decoder and the auxiliary decoder are combined to create pseudo labels, enabling the unlabeled data for supervision. For the inconsistent regions in predictions, we propose a self-contrast strategy that further improves the performance by reducing the discrepancy of the dual decoder for the same pixel. We conducted extensive experiments on three publicly available datasets and verified that our approach outperforms other methods for boundary quality. Specifically, with 5% labeled data on Left Atrial (LA) dataset, our proposed approach achieved a Boundary IoU 3.76% higher than the state-of-the-art methods. Code is available at https://github.com/SmileJET/CAC4SSL.