Cardiovascular AI: Multimodal Imaging (IVUS/OCT & Angiography) & Agentic Decision Support

At the Center for Digital Cardiovascular Innovations, University of Miami Miller School of Medicine / UHealth System (directed by Dr. Yiannis S. Chatzizisis, Chief, Division of Cardiovascular Medicine), this research develops multimodal deep learning and agentic clinical decision support systems (CDSS) to empower interventional cardiologists before and during complex percutaneous coronary interventions (PCI).

Multimodal Image Analysis: IVUS, OCT & Angiography

Effective clinical decision-making requires cross-scale structural awareness—from macroscopic vascular architecture to microscopic plaque vulnerability:

• High-Definition IVUS (HD-IVUS & NIRS-IVUS): Deep multi-task ConvNeXt-U-Net networks segmenting lumen, vessel wall (external elastic membrane), and plaque composition (calcium, fibrous, fibrolipidic) across 40,000+ expert-annotated IVUS frames.
• Intracoronary Optical Coherence Tomography (OCT): High-resolution (10–15 $\mu$m) optical profiling capturing thin-cap fibroatheromas (TCFA), macrophage infiltration, and acute post-stent malapposition.
• X-Ray Coronary Angiography (Luminography & QCA): Macroscopic vessel roadmapping, bifurcation anatomy tracking, and automated co-registration with pullback cross-sections.

Agentic Clinical Decision Support System (CDSS)

By orchestrating autonomous multi-agent reasoning over co-registered IVUS, OCT, and Angiographic streams, the CDSS moves beyond passive segmentation toward active procedural guidance:

• Automated Calcium Phenotyping: Quantitative assessment of calcium arc, longitudinal length, and radial depth to compute standardized calcium fracture risk scores.
• Agentic Procedural Triage: Autonomously recommends optimal lesion preparation strategies—selecting among non-compliant balloons, intravascular lithotripsy (IVL), rotational/orbital atherectomy, or cutting balloons based on individualized lesion mechanics.
• Virtual Stenting & Stent Under-Expansion Prediction: Integrates finite element analysis (FEA) and learned surrogate models across 1,200 lesion phenotypes to forecast stent expansion and mitigate adverse cardiac events (restenosis, thrombosis).

Related Manuscripts & Clinical Studies

• Deep Learning Model for Multi-Class Segmentation of High-Definition Intravascular Ultrasound — Under review at Scientific Reports (2026)
• Head-to-Head Comparison of Coronary Artery Bifurcation Stenting Strategies: A Virtual Clinical Study — Under review at JACC: Cardiovascular Interventions (2026)
• Head-to-Head Comparison of Latest Intracoronary Imaging Modalities — Submitted to JACC: Cardiovascular Interventions (2026)