# MG-HGLNet: A Mixed-Grained Hierarchical Geometric-Semantic Learning Framework with Dynamic Prototypes for Coronary Artery Lesions Assessment

## TL;DR

This paper introduces MG-HGLNet, a new deep learning framework for assessing coronary artery lesions using CCTA, which improves accuracy and works well with limited detailed labels.

## Contribution

MG-HGLNet introduces a novel mixed-grained learning framework with dynamic prototypes for coronary artery lesion assessment.

## Key findings

- MG-HGLNet achieves 92.4% stenosis grading accuracy and 91.5% plaque classification accuracy.
- The framework outperforms existing methods and works effectively under weakly supervised conditions.
- The proposed modules enhance global context modeling and plaque geometry-texture decoupling.

## Abstract

Automated assessment of coronary artery (CA) lesions via Coronary Computed Tomography Angiography (CCTA) is essential for the diagnosis of coronary artery disease (CAD). However, current deep learning approaches confront several challenges, primarily regarding the modeling of long-range anatomical dependencies, the effective decoupling of plaque texture from stenosis geometry, and the utilization of clinically prevalent mixed-grained annotations. To address these challenges, we propose a novel mixed-grained hierarchical geometric-semantic learning network (MG-HGLNet). Specifically, we introduce a topology-aware dual-stream encoding (TDE) module, which incorporates a bidirectional vessel Mamba (BiV-Mamba) encoder to capture global hemodynamic contexts and rectify spatial distortions inherent in curved planar reformation (CPR). Furthermore, a synergistic spectral–morphological decoupling (SSD) module is designed to disentangle task-specific features; it utilizes frequency-domain analysis to extract plaque spectral fingerprints while employing a texture-guided deformable attention mechanism to refine luminal boundary. To mitigate the scarcity of fine-grained labels, we implement a mixed-grained supervision optimization (MSO) strategy, utilizing anatomy-aware dynamic prototypes and logical consistency constraints to effectively leverage coarse branch-level labels. Extensive experiments on an in-house dataset demonstrate that MG-HGLNet achieves a stenosis grading accuracy of 92.4% and a plaque classification accuracy of 91.5%. The results suggest that our framework not only outperforms state-of-the-art methods but also maintains robust performance under weakly supervised settings, offering a promising solution for label-efficient CAD diagnosis.