Balancing Multi-Target Semi-Supervised Medical Image Segmentation With Collaborative Generalist and Specialists.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Current semi-supervised models struggle with simultaneous multi-target segmentation.
• Imbalanced target scales cause large targets to dominate loss, leading to misclassification of smaller targets.

Purpose of the Study:

• To address the performance decrease in simultaneous multi-target segmentation caused by scale imbalance.
• To propose a novel method, Collaborative Generalist and Specialists (CGS), for improved multi-target segmentation.

Main Methods:

• The CGS method employs a generalist for overall segmentation and dedicated specialists for each target class.
• Cross-consistency losses are developed to promote collaborative learning between the generalist and specialists.
• An inter-head error detection module is introduced to enhance pseudo-label quality.

Main Results:

• The CGS method achieves more balanced training, mitigating the dominance of large targets.
• Experimental results on three benchmarks demonstrate superior performance compared to state-of-the-art methods.
• The proposed approach effectively handles simultaneous segmentation of multiple medical targets.

Conclusions:

• The CGS method offers a significant advancement in semi-supervised multi-target medical image segmentation.
• The specialist-based approach effectively addresses the challenge of imbalanced target scales.
• This work provides a robust solution for accurate and balanced segmentation of multiple medical targets.