POPAR: Patch Order Prediction and Appearance Recovery for Self-supervised Medical Image Analysis

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Vision transformer-based self-supervised learning (SSL) shows promise for visual representation learning from unannotated images.
• Adaptation of SSL to medical imaging is limited due to domain discrepancies between photographic and medical images.

Purpose of the Study:

• Introduce POPAR (patch order prediction and appearance recovery), a novel vision transformer-based SSL framework tailored for chest X-ray images.
• Enable simultaneous learning of high-level contextual and fine-grained features for improved medical image analysis.

Main Methods:

• POPAR utilizes a vision transformer backbone to learn from chest X-ray images.
• The framework incorporates two tasks: correcting shuffled patch orders for contextual understanding and recovering patch appearance for fine-grained details.
• Pretrained POPAR models are evaluated on diverse downstream medical imaging tasks.

Main Results:

• POPAR demonstrates superior performance compared to state-of-the-art (SoTA) self-supervised models with vision transformer backbones.
• POPAR significantly outperforms three SoTA contrastive learning methods.
• The proposed method also achieves better results than fully-supervised pretrained models across various architectures.
• Ablation studies confirm the importance of both fine-grained and global contextual features for medical imaging tasks.

Conclusions:

• POPAR offers an effective self-supervised learning strategy for medical imaging, particularly chest X-rays.
• The framework's ability to capture both contextual and fine-grained features is crucial for enhancing performance on downstream tasks.
• The study provides a valuable resource with publicly available code and models for the research community.