ADNet++: A few-shot learning framework for multi-class medical image volume segmentation with uncertainty-guided...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep segmentation models require extensive labeled data, posing a barrier in the medical domain.
• Prototypical few-shot segmentation (FSS) models offer data-efficient alternatives but have limitations.

Purpose of the Study:

• To address shortcomings in prototypical FSS models: lack of uncertainty estimation, poor edge localization, and inability for one-step multi-class segmentation.
• To propose a novel methodology that enhances FSS model performance without requiring specific backbone architectures.

Main Methods:

• Developed a modified prototype extraction module for uncertainty map computation in FSS models.
• Introduced a novel feature refinement module guided by structural information and uncertainty maps for improved segmentation.
• Proposed a procedure for one-step multi-class FSS to avoid ambiguous predictions.

Main Results:

• Demonstrated that computed uncertainty maps effectively indicate model uncertainty.
• Showcased significant improvements in segmentation accuracy across multiple medical datasets (abdominal organ and cardiac segmentation).
• Achieved notable increases in Dice scores: +5.2% (CHAOS), +5.1% (BTCV), and +2.8% (MS-CMRSeg).

Conclusions:

• The proposed methodology enhances prototypical FSS models by incorporating uncertainty estimation, guided feature refinement, and efficient multi-class segmentation.
• The advancements offer a more robust and data-efficient solution for medical image segmentation tasks.
• The improvements are validated through rigorous evaluation on diverse medical imaging datasets.