Dual-Filter Cross Attention and Onion Pooling Network for Enhanced Few-Shot Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Few-shot learning shows promise in medical image segmentation.
• Current few-shot medical image segmentation (FSMIS) models have limitations in utilizing query information, leading to prototype bias and poor generalization.

Purpose of the Study:

• To propose the dual-filter cross attention and onion pooling network (DCOP-Net) for addressing limitations in FSMIS.
• To enhance prototype learning and segmentation accuracy in few-shot scenarios.

Main Methods:

• Introduced a dual-filter cross attention (DFCA) module to refine prototype learning by integrating query foreground features and avoiding feature entanglement.
• Designed an onion pooling (OP) module using eroding mask operations and masked average pooling to generate diverse prototypes and reduce bias.
• Implemented a parallel threshold perception (PTP) module and query self-reference regularization (QSR) strategy for improved segmentation robustness and consistency.

Main Results:

• DCOP-Net demonstrated superior performance compared to state-of-the-art methods on three public medical image datasets.
• The proposed modules effectively mitigated prototype bias and improved the integration of query image information.
• The network exhibited enhanced segmentation and generalization capabilities in few-shot learning tasks.

Conclusions:

• DCOP-Net offers a significant advancement in few-shot medical image segmentation.
• The novel DFCA and OP modules are effective in improving prototype learning and reducing bias.
• The proposed approach shows strong potential for clinical applications requiring accurate segmentation with limited data.