Semi-Supervised Unpaired Medical Image Segmentation Through Task-Affinity Consistency.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning-based semi-supervised learning (SSL) reduces manual annotation costs in medical image segmentation.
• Existing SSL methods often ignore shared information between labeled and unlabeled data, limiting performance.

Purpose of the Study:

• To develop a novel SSL approach for medical image segmentation that leverages shared information between labeled and unlabeled data.
• To improve the efficiency and accuracy of medical image segmentation by enhancing knowledge transfer.

Main Methods:

• Introduced a class-specific representation extraction approach with a task-affinity module.
• Utilized two views of feature maps (low-level context and structural information).
• Formulated a task-affinity consistency loss based on multi-scale class-specific representations.

Main Results:

• The proposed method significantly improved performance on three medical image segmentation datasets.
• Consistently outperformed existing state-of-the-art semi-supervised learning methods.
• Demonstrated the effectiveness of class-specific knowledge consistency for segmentation.

Conclusions:

• The novel SSL approach effectively mines shared information for improved medical image segmentation.
• Class-specific knowledge consistency is a promising strategy for enhancing SSL in medical imaging.
• The method offers a cost-effective solution for developing accurate medical image segmentation tools.