SELF-SUPERVISED CONTRASTIVE LEARNING FOR MEDICAL IMAGE SEGMENTATION WITH LIMITED LABELED DATA

Abstract

This study proposes a self-supervised contrastive learning framework for medical image segmentation under limited labeled-data conditions. The method first learns anatomical representations from a large pool of unlabeled medical images through contrastive pre-training, then fine-tunes a segmentation network using only a small annotated subset. The proposed pipeline is designed to improve label efficiency, reduce overfitting, and enhance boundary accuracy in challenging segmentation tasks. Experimental results show that the model consistently outperforms a fully supervised baseline across multiple annotation budgets, with the largest gains observed when labeled data are extremely scarce. The method also improves robustness, stability, and qualitative mask quality, particularly in low-contrast and small-structure cases. These findings indicate that unlabeled scans contain valuable structural information that can be effectively exploited for downstream segmentation. The study supports the use of self-supervised contrastive learning as a practical and scalable solution for clinical imaging environments where manual annotation is costly and limited.