Self-Supervised Interactive Embedding for One-Shot Organ Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• One-shot organ segmentation (OS2) is crucial for medical image analysis, requiring minimal annotation data.
• Existing OS2 methods struggle with slice similarity, necessitating complex allocation mechanisms.
• Exploring mutual information between reference (support) and test (query) slices is a core challenge.

Purpose of the Study:

• To develop a novel approach for one-shot organ segmentation that overcomes limitations of existing methods.
• To enhance the exploration of mutual information between support and query slices in medical imaging.
• To improve segmentation performance by effectively handling low-similarity slices.

Main Methods:

• Introduction of a novel Support-Query Interactive Embedding (SQIE) module.
• SQIE incorporates channel-wise co-attention, spatial-wise co-attention, and spatial bias transformation.
• Development of a self-supervised contrastive learning framework for feature embedding.

Main Results:

• The proposed SQIE module effectively mines interactive information between support and query slices.
• The method establishes feature connections even in slices with low similarity.
• Experiments on two large benchmarks demonstrate superior performance compared to existing methods.

Conclusions:

• The novel SQIE module significantly advances one-shot organ segmentation capabilities.
• The self-supervised contrastive learning framework enhances feature embedding for improved segmentation.
• The approach offers a powerful solution for medical image analysis with minimal annotation requirements.