From Few to More: Scribble-Based Medical Image Segmentation via Masked Context Modeling and Continuous Pseudo Labels.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Weakly supervised segmentation methods reduce annotation costs in medical imaging.
• Existing methods struggle with sparse annotations and rely on hard pseudo labels.
• Effective handling of varying annotation richness is crucial for model performance.

Purpose of the Study:

• To introduce MaCo, a novel weakly supervised model for medical image segmentation.
• To address limitations of current methods in handling sparse annotations.
• To develop a model that effectively learns from limited, scribble-based annotations.

Main Methods:

• MaCo utilizes Masked Context Modeling (MCM) with an attention-based masking strategy.
• Continuous Pseudo Labels (CPL) are generated using an exponential decay function on distance maps.
• The model learns from confidence maps derived from CPL, avoiding hard pseudo labels.

Main Results:

• MaCo demonstrated superior performance compared to existing weakly supervised methods.
• The model achieved state-of-the-art results across three public medical image segmentation datasets.
• The proposed MCM and CPL techniques effectively handle sparse annotation challenges.

Conclusions:

• MaCo offers a significant advancement in weakly supervised medical image segmentation.
• The "from few to more" principle, enabled by MCM and CPL, enhances model robustness.
• This work sets a new benchmark for scribble-based medical image segmentation.