Graph-Based Surgical Instrument Adaptive Segmentation via Domain-Common Knowledge.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) is crucial for surgical instrument segmentation, adapting models to new, unlabeled domains.
• Existing UDA methods often overlook domain-common knowledge, hindering their ability to capture inter-category relationships and leading to suboptimal performance.

Purpose of the Study:

• To propose a novel graph-based unsupervised domain adaptation framework, Interactive Graph Network (IGNet), for surgical instrument segmentation.
• To address the limitations of current UDA methods by effectively utilizing domain-common knowledge and inter-category relationships.

Main Methods:

• Developed the Domain-common Prototype Constructor (DPC) to aggregate feature maps into domain-common prototypes and construct a prototypical graph, capturing co-occurrent and long-range relationships.
• Introduced the Domain-common Knowledge Incorporator (DKI) using common-knowledge guidance and category-attentive graph reasoning to align feature maps across domains.
• Designed the Cross-category Mismatch Estimator (CME) for graph-based category-level alignment and pixel-wise adversarial weighting to refine feature distribution.

Main Results:

• Demonstrated the feasibility and superiority of IGNet through extensive experiments on three surgical instrument segmentation tasks.
• Achieved state-of-the-art performance compared to existing UDA methods.
• Ablation studies confirmed the effectiveness of individual components within the IGNet framework.

Conclusions:

• IGNet provides a robust solution for unsupervised domain adaptation in surgical instrument segmentation by effectively incorporating domain-common knowledge.
• The proposed graph-based approach significantly enhances model adaptation to unlabeled target domains, outperforming previous methods.