A Unified Visual Information Preservation Framework for Self-supervised Pre-Training in Medical Image Analysis

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Current self-supervised learning (SSL) methods in computer vision focus on comparative learning, which captures high-level semantics but lacks crucial local information for medical imaging tasks.
• This limitation hinders performance in applications like medical image-based diagnosis and tumor segmentation.

Purpose of the Study:

• To address the locality problem in comparative SSL by integrating pixel restoration for enhanced pixel-level information encoding.
• To incorporate scale information preservation, a valuable but underutilized feature in SSL for improved image understanding.

Main Methods:

• Developed a unified SSL framework (PCRLv2) using multi-task optimization on a feature pyramid.
• Implemented multi-scale pixel restoration and siamese feature comparison within the pyramid.
• Introduced non-skip U-Net for feature pyramid construction and sub-crop for 3D medical imaging.

Main Results:

• PCRLv2 demonstrated superior performance compared to existing self-supervised methods across diverse medical imaging tasks.
• Significant improvements were observed in brain tumor segmentation (BraTS 2018), chest pathology identification (ChestX-ray, CheXpert), pulmonary nodule detection (LUNA), and abdominal organ segmentation (LiTS).
• The framework achieved high performance even with limited annotations, outperforming counterparts by substantial margins in several cases.

Conclusions:

• PCRLv2 effectively enhances self-supervised learning for medical imaging by integrating pixel restoration and scale preservation.
• The proposed framework offers a powerful solution for medical image analysis, achieving state-of-the-art results with improved efficiency.