POPAR: Patch Order Prediction and Appearance Recovery for self-supervised learning in chest radiography.

Abstract

Self-supervised learning (SSL) has proven effective in reducing the dependency on large annotated datasets while achieving state-of-the-art (SoTA) performance in computer vision. However, its adoption in medical imaging remains slow due to fundamental differences between photographic and medical images. To address this, we propose POPAR (Patch Order Prediction and Appearance Recovery), a novel SSL framework tailored for medical image analysis, particularly chest X-ray interpretation. POPAR introduces two key learning strategies: (1) Patch order prediction, which helps the model learn anatomical structures and spatial relationships by predicting the arrangement of shuffled patches, and (2) Patch appearance recovery, which reconstructs fine-grained details to enhance texture-based feature learning. Using a Swin Transformer backbone, POPAR is pretrained on a large-scale dataset and extensively evaluated across multiple tasks, outperforming both SSL and fully supervised SoTA models in classification, segmentation, anatomical understanding, bias robustness, and data efficiency. Our findings highlight POPAR's scalability, strong generalization, and effectiveness in medical imaging applications. All code and models are available at GitHub.com/JLiangLab/POPAR (Version 2).