Few-Shot Medical Image Segmentation via Generating Multiple Representative Descriptors.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning advances medical image segmentation but requires large datasets.
• Data acquisition and annotation for medical images are challenging due to time, expertise, and privacy concerns.
• Few-shot Medical Image Segmentation (FSMIS) emerges as a critical research area to address data scarcity.

Purpose of the Study:

• To overcome the limitations of single-prototype representations in conventional FSMIS methods.
• To develop a more comprehensive approach for representing class distributions in medical images.
• To enhance the performance of few-shot medical image segmentation models.

Main Methods:

• Proposed Generate Multiple Representative Descriptors (GMRD) to capture class commonalities.
• Introduced a Multiple Affinity Maps based Prediction (MAMP) module for descriptor fusion.
• Developed novel loss functions to address intra-class variation and improve descriptor representativeness.
• Implemented a dual-path network design to balance foreground and background feature differences.

Main Results:

• The proposed GMRD method significantly outperforms existing state-of-the-art FSMIS techniques.
• Experimental validation on four public medical image datasets confirms the method's effectiveness.
• Ablation studies demonstrate the positive impact of the designed GMRD and MAMP modules.

Conclusions:

• The proposed approach effectively addresses the challenges of few-shot medical image segmentation.
• Generating multiple representative descriptors enhances the model's ability to segment medical images with limited data.
• The developed method offers a promising solution for practical applications in medical image analysis.