Self-Supervised Pre-Training Based On Contrastive Complementary Masking for Semi-Supervised Cardiac Image Segmentation
Yubo Zhou, Ran Gu, Shaoting Zhang, Guotai Wang · 2024
Cardiac structure segmentation is important for heart disease diagnosis, and deep learning with a large number of annotations has obtained remarkable performance on this task. Semi-Supervised Learning (SSL) has the potential to reduce annotation costs. However, most SSL methods only leverage unlabeled images for consistency regularization or pseudo labels, ignoring their potential for feature learning with pretraining. In this work, we propose a novel framework that utilizes self-supervised pre-training for better semi-supervised segmentation. Our framework consists of two modules: 1) Self-supervised pre-training based on Contrastive Complementary Masking (CCM), where a contrastive loss is used for two networks that encode complimentary masked versions of the same input, in addition to a reconstruction loss to enhance global and local feature learning; 2) Semi-supervised segmentation with Cross Pseudo Supervision (CPS) between the two pre-trained networks, where each network is supervised by pseudo labels from the other to deal with unlabeled images. Experiments on the ACDC dataset showed that our method improved performance by 6.73 percentage points over baseline with a 5% annotation ratio, and outperformed three state-of-the-art semi-supervised methods.