CR-GLoCo: Cross-Resolution Learning via Global-Local Context Consistency for semi-supervised 3D medical segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels at 3D medical image segmentation but requires extensive expert annotations.
• Voxel-level annotation is time-consuming and expensive, motivating the use of semi-supervised learning (SSL).
• Existing SSL methods often use patch/slice-based training, losing global context and hindering generalization.

Purpose of the Study:

• To develop an efficient semi-supervised learning framework for 3D medical image segmentation.
• To address the loss of global anatomical context in patch/slice-based SSL methods.
• To reduce the annotation burden in medical image segmentation.

Main Methods:

• Proposed CR-GLoCo, a cross-resolution learning framework enforcing Global-Local Context Consistency.
• Coupled a low-resolution global branch with a high-resolution local branch for holistic priors and fine boundaries.
• Implemented mutual pseudo-supervision with confidence filtering and overlap-based sampling for robust context transfer.

Main Results:

• CR-GLoCo achieved superior performance compared to state-of-the-art SSL methods on three challenging datasets.
• The framework effectively leverages both global anatomical information and local details.
• Demonstrated improved generalization under scarce labeled data conditions.

Conclusions:

• CR-GLoCo offers a powerful solution for semi-supervised 3D medical image segmentation.
• The cross-resolution approach enhances context utilization and model robustness.
• The method significantly alleviates the need for large annotated datasets in medical imaging AI.