IGU-Aug: Information-Guided Unsupervised Augmentation and Pixel-Wise Contrastive Learning for Medical Image Analysis

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Contrastive learning (CL) is a key self-supervised learning technique.
• Instance-level CL is well-researched, but pixel-wise CL lacks tailored augmentation strategies.
• Pixel-wise CL is crucial for dense prediction tasks.

Purpose of the Study:

• To develop a novel pixel augmentation method for pixel-wise contrastive learning.
• To enhance feature representation at a pixel granularity.
• To improve performance in unsupervised local feature matching.

Main Methods:

• Classified pixels into low-, medium-, and high-informative categories based on information content.
• Designed adaptive augmentation strategies (intensity, sampling ratio) for each pixel category.
• Validated the approach through extensive experiments.

Main Results:

• The proposed information-guided pixel augmentation strategy encodes more discriminative representations.
• Achieved superior performance in unsupervised local feature matching compared to existing methods.
• The pretrained model boosted performance for both one-shot and fully supervised models.

Conclusions:

• This work presents the first pixel augmentation method with pixel granularity for unsupervised pixel-wise contrastive learning.
• The adaptive strategy effectively enhances feature learning for dense prediction tasks.
• The approach offers a significant advancement in self-supervised learning for pixel-level analysis.