MG-HGLNet: A Mixed-Grained Hierarchical Geometric-Semantic Learning Framework with Dynamic Prototypes for Coronary...

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Automated assessment of coronary artery lesions using Coronary Computed Tomography Angiography (CCTA) is crucial for diagnosing coronary artery disease (CAD).
• Existing deep learning methods struggle with long-range dependencies, plaque texture vs. stenosis geometry decoupling, and mixed-grained annotations.
• There is a need for advanced models that can effectively handle these challenges for improved CAD diagnosis.

Purpose of the Study:

• To introduce a novel mixed-grained hierarchical geometric-semantic learning network (MG-HGLNet) for automated CA lesion assessment.
• To address the limitations of current deep learning approaches in modeling anatomical dependencies, feature decoupling, and utilizing mixed-grained annotations.
• To develop a robust and label-efficient framework for CAD diagnosis.

Main Methods:

• Proposed a topology-aware dual-stream encoding (TDE) module with a bidirectional vessel Mamba (BiV-Mamba) encoder for global context and spatial distortion correction.
• Introduced a synergistic spectral-morphological decoupling (SSD) module using frequency analysis and texture-guided attention to separate plaque spectral and geometric features.
• Implemented a mixed-grained supervision optimization (MSO) strategy leveraging anatomy-aware prototypes and logical constraints for coarse-grained labels.

Main Results:

• MG-HGLNet achieved a stenosis grading accuracy of 92.4% and a plaque classification accuracy of 91.5% on an in-house dataset.
• The proposed framework demonstrated superior performance compared to state-of-the-art methods.
• The model maintained robust performance in weakly supervised settings, indicating its effectiveness with limited fine-grained labels.

Conclusions:

• MG-HGLNet offers a promising solution for label-efficient CAD diagnosis by effectively addressing challenges in automated CA lesion assessment.
• The network's ability to model long-range dependencies, decouple features, and utilize mixed-grained annotations represents a significant advancement.
• The findings suggest a potential for improved clinical workflows in CAD diagnosis through advanced AI techniques.