Effective Semi-Supervised Medical Image Segmentation With Probabilistic Representations and Prototype Learning.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised medical image segmentation faces challenges like label scarcity, class imbalance, and data uncertainty.
• Data uncertainty, particularly at the pixel level, is a critical yet often overlooked issue in existing methods.

Purpose of the Study:

• To propose a novel probabilistic prototype-based classifier to explicitly model and address data uncertainty in medical image segmentation.
• To enhance model robustness against ambiguous boundaries and noise by incorporating uncertainty estimation throughout the classification process.

Main Methods:

• Developed a probabilistic prototype-based classifier integrating uncertainty estimation into representation formulation, pixel-prototype matching, and prototype updates.
• Leveraged principles from probability theory for a comprehensive uncertainty-aware approach.
• Evaluated the framework on three public datasets with significant boundary ambiguity and simulated noisy data.

Main Results:

• The proposed method significantly enhances model robustness to tricky pixels, including ambiguous boundaries and noise, compared to deterministic and other uncertainty-aware strategies.
• Empirical evaluations demonstrated the superiority of the probabilistic approach over several competing methods on challenging datasets.
• The framework showed improved robustness when subjected to simulated noisy data.

Conclusions:

• Explicitly modeling data uncertainty at the pixel level is crucial for improving the robustness of semi-supervised medical image segmentation.
• The probabilistic prototype-based classifier offers a promising direction for handling data uncertainty and enhancing segmentation performance in complex medical imaging scenarios.