Cross-view discrepancy-dependency network for volumetric medical image segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Limited data is a major challenge for deep learning-based volumetric medical image segmentation.
• Existing multi-view slice methods often neglect inter-slice spatial continuity and cross-view relationships.

Purpose of the Study:

• To propose a novel network, Cross-View Discrepancy-Dependency Network (CvDd-Net), for volumetric medical image segmentation.
• To enhance volume representation learning by exploiting multi-view slice priors, specifically addressing view discrepancy and dependency.

Main Methods:

• Developed a discrepancy-aware morphology reinforcement (DaMR) module to learn view-specific representations using morphological information (object boundary and position).
• Designed a dependency-aware information aggregation (DaIA) module to integrate multi-view slice information based on cross-view dependency.
• Evaluated the method on fully-supervised and semi-supervised tasks using four medical image datasets: Thyroid, Cervix, Pancreas, and Glioma.

Main Results:

• The proposed CvDd-Net effectively improves volumetric medical image segmentation performance.
• The DaMR and DaIA modules successfully leverage multi-view information by addressing discrepancies and dependencies.
• Demonstrated significant efficacy across diverse medical imaging datasets and segmentation tasks.

Conclusions:

• CvDd-Net offers a robust solution for volumetric medical image segmentation, particularly in data-limited scenarios.
• Exploiting cross-view discrepancy and dependency is crucial for advancing multi-view representation learning in medical imaging.