Semantic-Aware Contrastive Learning for Multi-Object Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial but challenging.
• Contrastive learning can improve neural network performance without ground-truth labels.
• Traditional contrastive learning struggles with multi-object segmentation in medical images.

Purpose of the Study:

• To develop a semantic-aware contrastive learning method for multi-object medical image segmentation.
• To adapt contrastive learning from image-level to pixel-level segmentation tasks.
• To improve the accuracy of medical image segmentation across diverse clinical cohorts.

Main Methods:

• Proposed a semantic-aware contrastive learning approach using attention masks and image-wise labels.
• Embedded different semantic objects into distinct clusters for improved segmentation.
• Evaluated the method on multi-organ segmentation tasks using in-house and public datasets (MICCAI Challenge 2015 BTCV, FLARE 2021).

Main Results:

• Achieved substantial improvements in Dice scores (5.53% and 6.09%) on medical image segmentation cohorts.
• Demonstrated significant performance gains on an external medical image cohort (Dice 0.922 to 0.933).
• Results showed statistically significant improvements (p-value 0.01) compared to state-of-the-art methods.

Conclusions:

• The proposed semantic-aware contrastive learning method effectively advances multi-object medical image segmentation.
• The approach enhances encoder-decoder neural network performance without requiring voxel-wise labels.
• This method offers a robust solution for improving segmentation accuracy in large-scale clinical applications.