An artificial intelligence-based model for exploration of coronary artery dissection and planning of percutaneous coronary intervention

Coronary artery dissections with the morphologic characteristics of false lumens (FL) are often encountered during percutaneous coronary Intervention (PCI) procedure, but there lack effective approaches to provide explorative guidance.

We aimed to propose a novel machine learning framework for reducing complications through more accurate intervention guidance.

We developed a comprehensive artificial intelligence pipeline for precise detection and visualization of coronary dissections from digital subtraction angiography (DSA) images. The dissection detection was performed using the YOLOv8 algorithm, followed by segmentation with an SO(2)-Equivariant-UNet (EQUNet) model specifically designed for coronary vessel segmentation. In this multicentre retrospective cohort study, 131 patients with dissection and 180 patients without dissection were included.

For dissection segmentation, EQUNet achieved the highest precision of 87.51% and an F1 score of 83.03%, along with the third-highest recall (79.41%) and the second-highest Jaccard index (0.7120). In DSA image segmentation, EQUNet achieved the highest accuracy (98.03%) and specificity (87.51%) among all compared Transformer-based and CNN-based methods. Furthermore, the detection model demonstrated strong potential in accurately localizing the FL, achieving a confidence level of 0.869. Notably, in patients without dissections, no false positives were observed. The dissection detection algorithm achieved a high specificity of 99.7% and sensitivity of 95.8%, effectively distinguishing dissections from non-dissections, including calcified lesions.

The proposed advanced AI pipeline demonstrates the capability to precisely detect and localize coronary dissections from DSA images, followed by high-fidelity segmentation, ensuring detailed and clinically actionable visualization. It holds promise for providing tailored, patient-specific guidance in minimally invasive cardiac procedures.

Overall Framework of study design