HyperSeg-DG: multi-scale hyper feature context for domain-generalized medical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation faces challenges with domain shifts and ambiguous boundaries.
• Existing models struggle with generalization across different imaging modalities and scanners.
• Limited robustness in real-world clinical applications due to separate handling of domain shifts and boundary complexities.

Purpose of the Study:

• To develop a novel medical image segmentation approach for improved domain generalization and accuracy.
• To address foreground-background uncertainty and boundary ambiguities in medical images.
• To enhance the robustness of segmentation models in diverse and challenging clinical scenarios.

Main Methods:

• Proposed HyperSeg-DG, integrating the WMamba backbone with the Multi-Scale Hyper Feature Context Block (HFCB).
• HFCB captures multi-scale feature relations and long-range dependencies to resolve ambiguities.
• WMamba processes images in localized windows with selective 2D scanning for robust feature learning.

Main Results:

• HyperSeg-DG demonstrated consistent 2-3% improvements over strong baselines across multiple benchmarks.
• The model effectively focuses on relevant pathological features, reducing interference from irrelevant ones.
• Achieved enhanced segmentation performance and generalization across diverse, unseen domains.

Conclusions:

• HyperSeg-DG offers a robust solution for medical image segmentation, improving generalization.
• The integration of WMamba and HFCB effectively tackles domain shifts and boundary ambiguities.
• The proposed method shows significant potential for clinical deployment and real-world applications.