Anatomy-Aware Contrastive Pre-training: Leveraging Spatial Consistency for Label-Efficient Medical Image Diagnosis Across Multi-Modal Imaging

Abstract

Medical image analysis faces persistent challenges in acquiring expert annotations due to high costs and specialized expertise requirements. Self-supervised learning offers a promising solution by learning representations from unlabeled data. This paper introduces an anatomy-aware contrastive pre-training framework that exploits spatial consistency and anatomical structure priors inherent in medical images. The proposed approach integrates contrastive learning with anatomical constraints, enabling effective knowledge transfer across CT, MRI, and X-ray modalities. Through comprehensive experiments on multiple diagnostic tasks, the framework demonstrates superior label efficiency, achieving competitive performance with only 10% of labeled data compared to fully supervised baselines. The cross-modal evaluation reveals consistent improvements of 8.3% in classification accuracy and 6.7% in segmentation Dice scores. These results validate the effectiveness of incorporating anatomical priors into self-supervised learning pipelines for medical imaging applications.