Correspondence-based Generative Bayesian Deep Learning for semi-supervised volumetric medical image segmentation.

Area of Science:

• Medical image analysis
• Computer-aided diagnosis
• Machine learning in healthcare

Background:

• Automated medical image segmentation is vital for clinical applications.
• High annotation costs limit fully-supervised methods, driving interest in semi-supervised approaches.
• Existing semi-supervised methods struggle with pseudo-label quality due to data distribution bias.

Purpose of the Study:

• To introduce an innovative Correspondence-based Generative Bayesian Deep Learning (C-GBDL) model.
• To improve the generation of high-quality pseudo-labels in semi-supervised medical segmentation.
• To address data distribution bias and enhance segmentation accuracy.

Main Methods:

• Developed a teacher-student architecture incorporating a multi-scale semantic correspondence method.
• Teacher model learns generalized data distribution via feature matching with reference volumes.
• Proposed a double uncertainty estimation schema (predictive entropy and structural similarity) to refine pseudo-labels.

Main Results:

• The C-GBDL model demonstrated superior performance in comparative experiments.
• Effectiveness validated on two public medical datasets.
• Achieved improved segmentation accuracy compared to existing semi-supervised methods.

Conclusions:

• The proposed C-GBDL model effectively generates high-quality pseudo-labels for semi-supervised medical image segmentation.
• The multi-scale semantic correspondence and double uncertainty estimation significantly improve segmentation performance.
• This approach offers a promising solution to reduce annotation costs and enhance clinical utility.