Few-shot medical image segmentation with high-fidelity prototypes

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Few-shot semantic segmentation (FSS) models excel with distinct objects and simple backgrounds.
• Existing FSS methods are suboptimal for medical imaging due to complex backgrounds and less distinct objects.
• High-fidelity prototype representation is crucial for FSS in challenging medical scenarios.

Purpose of the Study:

• To propose a novel DetailSelf-refinedPrototypeNetwork (DSPNet) for accurate FSS in medical imaging.
• To develop a model capable of constructing comprehensive foreground and background prototypes.
• To overcome limitations of existing FSS methods in complex medical image analysis.

Main Methods:

• DSPNet constructs foreground prototypes by modeling multimodal structures via clustering and channel-wise fusion.
• Background prototypes are generated by integrating channel-specific structural information under sparse channel-aware regulation.
• The network focuses on maintaining global semantics while preserving detailed semantics for high-fidelity prototypes.

Main Results:

• DSPNet demonstrates superior performance over state-of-the-art methods on three challenging medical image benchmarks.
• The proposed method effectively handles complex backgrounds and subtle object details in medical images.
• Experimental results validate the efficacy of DSPNet in few-shot semantic segmentation tasks.

Conclusions:

• DSPNet offers a significant advancement for few-shot semantic segmentation in medical imaging.
• The model's ability to create high-fidelity prototypes addresses key limitations of prior FSS approaches.
• DSPNet provides a robust solution for segmenting complex structures in medical scans.