Weakly supervised learning for multi-class medical image segmentation via feature decomposition.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Weakly supervised learning reduces annotation effort for deep learning in medical image segmentation.
• Existing methods primarily focus on single-class segmentation, leaving multi-class segmentation underexplored.
• Medical images present unique challenges like label symbiosis and location adjacency in multi-class segmentation.

Purpose of the Study:

• To develop a novel weakly supervised learning method for multi-class medical image segmentation using image-level labels.
• To address challenges of label symbiosis and location adjacency in medical image segmentation.
• To improve the performance of automated medical image segmentation.

Main Methods:

• A multi-level classification network encodes multi-scale features for class-specific binary predictions and Class Activation Maps (CAMs).
• A feature decomposition module based on semantic affinity learns class-independent and class-dependent features to maximize inter-class feature distance.
• A cross-guidance loss and a mutually exclusive loss are employed to alleviate label symbiosis and minimize region overlap, respectively.

Main Results:

• The proposed framework demonstrates superior performance in both single-class and multi-class medical image segmentation across three datasets.
• The method effectively addresses label symbiosis and location adjacency challenges.
• Achieved state-of-the-art results in weakly supervised multi-class medical image segmentation.

Conclusions:

• The developed weakly supervised learning method offers a promising solution for multi-class medical image segmentation with image-level labels.
• The approach provides a foundation for future research in challenging multi-class segmentation tasks.
• The findings highlight the potential of weakly supervised learning to advance medical image analysis.