AI-Enabled Automatic Coronary Vessel Extraction

Automating coronary vessel segmentation to streamline clinical analysis.

Coronary vessel extraction from CT angiography is a prerequisite step for downstream plaque analysis, stenosis quantification, and fractional flow reserve (FFR) predictions. This project delivered a fully automated deep learning pipeline that minimizes the labor-intensive manual tracing, dramatically accelerating the analysis workflow.

The work spanned the full ML lifecycle: from data curation and annotation quality control, through model architecture selection and training, to post-processing refinement and production deployment on a distributed computing cluster.

From data curation to production inference.

High-quality training data is the foundation of getting accurate and robust inference results in production for the analysis workflow. A significant portion of this project was devoted to systematic data curation — sourcing, cleaning, and quality-grading coronary CTA scans — through collaborations with clinical analysts to segment coronary vessel lumen and annotate labels.

Model training utilized PyTorch on GPU clusters. Post-processing steps — topology correction, anatomical constraint enforcement, and centerline extraction — were critical to delivering vessel trees suitable for downstream quantitative analysis. A Directed Acyclic Graph (DAG) was designed to use with HTCondor in AWS to perform inference and image processing to compute plaque components.

• Data Curation High-fidelity training dataset through representative data selection across sites, scanners, and disease severity, combined with scan quality grading, annotation QC, and iterative label refinement
• Model Development The nnU-Net segmentation framework was chosen for this work. Training was performed using PyTorch on hardware with multiple GPU instances and model performance was assessed with lumen surface-to-surface distance and percentage vessel segment coverage against analyst segmentation
• Post-Processing Lumen refinement, centerline extraction, vessel labeling are essential to downstream plaque component analysis
• Deployment Initial product integration is based on PyTorch model. A scalable deployment version using TensorRT model and served via NVIDIA Triton Inference Server is ready for integration