Multi-level Asymmetric Contrastive Learning for Medical Image Segmentation Pre-training

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation is crucial but hindered by the difficulty of obtaining expert-labeled data.
• Existing contrastive learning methods for medical images often overlook multi-level representations and underutilize decoders.

Purpose of the Study:

• To propose a novel multi-level asymmetric contrastive learning framework (MACL) to enhance medical image segmentation.
• To address limitations of current contrastive learning by enabling simultaneous encoder-decoder pre-training and integrating multi-level representations.

Main Methods:

• Developed an asymmetric contrastive learning structure for simultaneous encoder and decoder pre-training.
• Implemented a multi-level contrastive strategy integrating feature-level, image-level, and pixel-level correspondences.
• Evaluated the framework on 8 medical image datasets against 11 existing contrastive learning strategies.

Main Results:

• MACL demonstrated superior performance compared to 11 other contrastive learning strategies.
• Achieved significant Dice score improvements (1.72% to 7.87%) on datasets like ACDC, MMWHS, HVSMR, and CHAOS with only 10% labeled data.
• Showcased strong generalization capabilities across 5 different U-Net backbones.

Conclusions:

• The proposed MACL framework effectively enhances medical image segmentation by leveraging multi-level representations and simultaneous encoder-decoder pre-training.
• MACL offers a promising solution for improving segmentation accuracy, particularly in low-data regimes, and exhibits robust generalization.