SuperCL: Superpixel Guided 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 limited expert-annotated datasets.
• Existing contrastive learning methods often overlook intra-image pixel relationships and rely on inefficient manual thresholding for pair generation.

Purpose of the Study:

• To introduce SuperCL, a novel contrastive learning approach for medical image segmentation pre-training.
• To address limitations in current methods by exploiting structural priors and pixel correlations.

Main Methods:

• SuperCL utilizes superpixel maps for generating pseudo masks to guide supervised contrastive learning.
• Introduces Intra-image Local Contrastive Pairs (ILCP) and Inter-image Global Contrastive Pairs (IGCP) generation strategies.
• Incorporates Average SuperPixel Feature Map Generation (ASP) and Connected Components Label Generation (CCL) modules for enhanced structural information exploitation.

Main Results:

• SuperCL demonstrated superior performance across 8 medical image datasets compared to 12 existing methods.
• Achieved significant improvements in Dice Similarity Coefficient (DSC), including 3.15%, 5.44%, and 7.89% higher scores on MMWHS, CHAOS, and Spleen datasets, respectively, with only 10% annotations.
• Visualization figures confirmed more precise segmentation predictions.

Conclusions:

• SuperCL offers an effective solution for medical image segmentation pre-training, particularly in low-data regimes.
• The proposed contrastive pair generation strategies and structural information exploitation modules significantly advance the state-of-the-art.
• The open-sourced code facilitates further research and application in medical image analysis.