Interactive prototype learning and self-learning for few-shot medical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Few-shot learning (FSL) reduces the need for extensive labeled data in medical image segmentation.
• Traditional deep learning methods face performance gaps in FSL, especially with distribution shifts between support and query images.
• Key challenges include intra-class inconsistency and inter-class similarity, leading to blurred segmentation boundaries.

Purpose of the Study:

• To develop an advanced network for few-shot medical image segmentation.
• To overcome limitations of existing FSL methods in handling distribution differences.
• To improve segmentation accuracy and generalization ability for unseen medical imaging tasks.

Main Methods:

• Proposed a deep encoding-decoding module for high-level feature extraction and peak prototype generation.
• Introduced an interactive prototype learning module for feature consistency and similarity reduction via mean and peak prototype interactions.
• Implemented a query features-guided self-learning module to refine segmentation and incorporate boundary details from low-level features.

Main Results:

• The proposed model achieved competitive segmentation performance on benchmark datasets.
• Demonstrated substantial improvements in generalization ability for new segmentation tasks.
• Effectively addressed intra-class inconsistency and inter-class similarity challenges.

Conclusions:

• The interactive prototype and self-learning network offers a robust solution for few-shot medical image segmentation.
• The method enhances feature representation and boundary definition, leading to superior performance.
• This approach significantly improves the adaptability of segmentation models to novel tasks.