Related Experiment Video
Updated: Aug 5, 2026
Semi-Automatic Graphical Tool for Measuring Coronary Artery Spatially Weighted Calcium Score from Gated Cardiac Computed Tomography Images
Published on: September 22, 2023
Preoperative Decision-Oriented Decoupled Multi-Scale Feature Pyramid and Global Context Aggregation Network for
Jianhua Lin1, Lin Lin1, Zhenzhen Li1
1Fuqing City Hospital Affiliated to Fujian Medical University, Fuqing, China.
Insights
This study introduces a new AI model for precise coronary artery segmentation in X-ray angiography, improving cardiovascular diagnosis and treatment planning by overcoming common segmentation challenges.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiovascular Imaging
Background:
- Accurate preoperative coronary artery delineation is crucial for cardiovascular diagnosis and treatment.
- Challenges include class imbalance, poor bifurcation contrast, and overlapping artifacts, leading to segmentation errors.
Purpose of the Study:
- To develop an advanced AI framework for robust coronary artery segmentation.
- To improve the accuracy and reliability of vessel segmentation in X-ray angiography.
Main Methods:
- A decoupled hybrid architecture separating multi-scale detail recovery and global context aggregation.
- Utilizes a Feature Pyramid Network (FPN) neck for edge preservation and a UPerHead for contextual modeling.
Main Results:
- Achieved Dice, IoU, and Centreline Dice scores of 0.7633, 0.6290, and 0.7827 on the ARCADE benchmark.
- Outperformed baseline methods and demonstrated improved contextual attention and interpretability.
Conclusions:
- The proposed framework enables robust preoperative vessel segmentation.
- Supports better assessment of distal continuity, bifurcation morphology, and lesion-adjacent boundaries for clinical decisions.
Background:
Accurate preoperative coronary artery delineation in X-ray angiography is critical for stenosis quantification, cardiovascular diagnosis, and treatment planning, yet class imbalance, weak bifurcation contrast, and overlapping artefacts often cause broken or distorted vessel segmentation.
Methods:
We propose a decoupled hybrid architecture that separates multi-scale detail recovery from global context aggregation, combining an FPN neck for distal-vessel edge preservation with a UPerHead head for multi-level contextual prior modelling.
Results:
On the ARCADE benchmark, the model achieves Dice, IoU, and Centreline Dice scores of 0.7633, 0.6290, and 0.7827, outperforming competitive baseline methods in terms of Dice, IoU, and Centreline Dice, with consistent but moderate improvements over the strongest competing model. Grad-CAM++ and feature-distribution analyses suggest improved contextual attention and enhanced interpretability of learnt representations.
Conclusions:
The framework enhances robust preoperative vessel segmentation, supporting assessment of distal continuity, bifurcation morphology, and lesion-adjacent boundaries for clinical decision-making.
