SSCLMix: A self-supervised contrastive learning-based data mixing augmentation method

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning (DL) for medical image segmentation struggles with limited and imbalanced data, hindering lesion feature learning and performance.
• Existing data mixing augmentation methods can degrade image structure and cause feature misalignment, impacting mixed sample quality.

Purpose of the Study:

• To propose a novel self-supervised contrastive learning-based image mixing method (SSCLMix) to address data scarcity and imbalance in medical image segmentation.
• To improve the quality of mixed samples and enhance the performance of segmentation models.

Main Methods:

• SSCLMix classifies training samples by structural similarity for targeted mixing.
• It employs dual-encoder contrastive learning and cross-self-attention for cross-sample modeling to generate mixed images.
• A dual-spatial feature perception residual module (DSFPR) is introduced to preserve image structure and regional information.

Main Results:

• SSCLMix generates higher-quality mixed samples compared to existing data augmentation methods.
• The proposed method significantly improves segmentation model metrics across seven medical image segmentation tasks.
• SSCLMix demonstrates competitive computational efficiency and practicality.

Conclusions:

• SSCLMix effectively overcomes data limitations in medical image segmentation by generating superior mixed samples.
• The method offers a promising approach to enhance DL model performance in medical image analysis.
• SSCLMix provides a practical and efficient solution for improving medical image segmentation accuracy.