Mask-Guided Proxy Mining Network for Few-Shot Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot medical image segmentation (FSMIS) methods rely on limited labeled data for new classes.
• Current FSMIS approaches use pixel-level correlations, which can cause mismatches and semantic ambiguity due to class information gaps.

Purpose of the Study:

• To introduce a novel Mask-Guided Proxy Mining Network (MPMNet) to address foreground-background ambiguity in FSMIS.
• To improve the accuracy and robustness of medical image segmentation with limited data.

Main Methods:

• MPMNet mines representative features (proxies) from support and query images to resolve ambiguity.
• A mask-guided proxy mining module adaptively learns proxies to handle variations in scale and shape.
• Hierarchical prior generation and context-aware feature enrichment modules enhance multi-scale information and feature discriminability.

Main Results:

• MPMNet effectively overcomes false pixel matches by establishing proxy-level semantic correlations.
• Experiments on three benchmarks show MPMNet significantly outperforms state-of-the-art methods.
• A mean gain of 2.71% in Dice Similarity Coefficient (DSC) was achieved across datasets.

Conclusions:

• MPMNet offers a robust solution for FSMIS by mitigating semantic ambiguity.
• The proposed network architecture enhances segmentation performance in low-data regimes.
• The approach demonstrates significant improvements in medical image segmentation accuracy.