Boundary-Aware Prototype in Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Semi-supervised medical image segmentation (SSMIS) methods often train labeled and unlabeled data separately, limiting the effective use of true labels.
• Existing approaches struggle to fully leverage the supervisory information from limited true labels in SSMIS.

Purpose of the Study:

• To propose a novel consistency learning framework for SSMIS that enables interactive training between labeled and unlabeled data.
• To maximize the utilization of true labels by bridging the gap between separate data training paradigms.

Main Methods:

• Developed a framework combining CNN-based linear classification and non-parametric nearest neighbor classification using boundary-aware prototypes.
• Prototypes are clustered from both labeled and unlabeled data features, acting as an interactive bridge for training.
• Introduced pixel-prototype contrastive learning to enhance feature discriminability for non-parametric distance measurement.

Main Results:

• The proposed method, using a lightweight UNet backbone, achieved superior performance compared to a 3D VNet with more parameters.
• Demonstrated effective utilization of true labels through interactive training facilitated by prototypes.
• Showcased improved feature discriminability and suitability for non-parametric distance measurement.

Conclusions:

• The novel consistency learning framework effectively integrates labeled and unlabeled data in SSMIS.
• Boundary-aware prototypes serve as a crucial mechanism for interactive training and enhanced label utilization.
• The method offers a promising direction for improving SSMIS performance with limited labeled data.