SACL: Semantics-Aware Contrastive Learning for One-Shot Medical Image Segmentation

One-shot medical image segmentation remains a challenging task due to the diversified anatomical structures and the presence of ambiguous surrounding tissues. While large pre-trained models have shown success in few-shot segmentation, they typically require extensive datasets for fine-tuning. Approaches such as synthetic pseudo-labeling and semi-supervised learning offer simpler, cost-effective solutions by learning from noisy annotations with consistency regularization. However, a significant challenge is the domain shift between homogenous synthetic data and more complex clinical data, which limits the generalizability of these models. In this paper, we propose a novel semantics-aware contrastive learning (SACL) method for one-shot medical image segmentation. SACL includes two key components: (1) semantic-aware image translation (SAIT), which addresses the domain shift between synthetic and clinical data, and (2) semantic-aware feature enhancement (SAFE), which leverages pseudo-labeling to enhance feature representation. Experimental results show that the proposed SACL approach achieves significant performance improvements over compared state-of-the-art methods. Code available at https://github.com/ltl5207/OSMASSEG.