Linear semantic transformation for semi-supervised medical image segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Deep learning significantly advances medical image segmentation for diagnosis and planning.
• Current methods struggle with semantic learning efficiency due to reliance on extensive annotations.
• Robust semantic representation in latent spaces remains a key challenge.

Purpose of the Study:

• To propose a novel semi-supervised learning framework for medical image segmentation.
• To address the inefficiency of semantic learning by reducing reliance on annotated data.
• To construct generalized representations from diverse semantics for improved segmentation.

Main Methods:

• Developed a self-supervised learning component for context recovery via image reconstruction (spatial and intensity).
• Employed linear semantic transformation to convert semantic-rich feature maps into image segmentation.
• Validated the framework on five diverse medical image segmentation datasets.

Main Results:

• Achieved top performance on IXI, ScaF, COVID-19-Seg, PC-Seg, and Brain-MR datasets with scores ranging from 47.50% to 73.78%.
• Outperformed state-of-the-art semi-supervised methods, achieving Dice Similarity Coefficient (DSC) values of 77.15% and 75.22% on representative datasets.
• Demonstrated the effectiveness, simplicity, and ease-of-use of linear semantic transformation in semi-supervised medical image segmentation.

Conclusions:

• The proposed framework successfully enhances medical image segmentation using semi-supervised learning.
• Linear semantic transformation is a simple yet powerful tool for achieving robust segmentation with limited annotations.
• The method offers a promising direction for developing intelligent medical systems with improved segmentation capabilities.