AI-driven coronary stenosis detection versus IVUS reference standard

Coronary artery disease (CAD) remains a major cause of cardiovascular-related mortality. Accurate detection of arterial stenosis is critical for guiding clinical decisions. While coronary angiography is the gold standard for diagnosing CAD, its manual interpretation is prone to subjectivity and variability. Intravascular Ultrasound (IVUS) provides high-resolution cross-sectional vessel imaging and serves as a valuable reference for evaluating stenosis severity.

This study aims to evaluate a deep learning-based framework for automatic segmentation and stenosis quantification from coronary angiograms, and to assess its performance by comparing it with IVUS-based assessments.

We use a hybrid deep learning model that integrates MedSAM and VM-UNet architectures to perform high-accuracy segmentation of coronary arteries in angiographic images. Post-segmentation, we extract the vascular centerline, compute vessel diameters, and measure the degree of stenosis. To evaluate the clinical reliability of the proposed method, its quantitative performance will be compared against IVUS-derived measurements, which serve as the reference standard for stenosis assessment

Using a mixed dataset (ARCADE, DCA1, and GH), the proposed model achieved an average IoU of 0.6308, with sensitivity and specificity of 0.9772 and 0.9903, respectively. On the ARCADE dataset alone, IoU reached 0.6303, with sensitivity of 0.9832 and specificity of 0.9933. The stenosis detection component demonstrated a true positive rate (TPR) of 0.5867 and a positive predictive value (PPV) of 0.5911. Comparative analysis with IVUS data is underway to evaluate concordance and potential clinical applicability.

The SAM-VMNet model demonstrates robust performance in segmenting coronary arteries and detecting stenosis from angiograms. By benchmarking against IVUS, we aim to validate the model's diagnostic accuracy and its potential as a non-invasive, cost-effective tool for CAD assessment.