Biomechanics-trained diffusion model predicts post-PCI plaque morphology and rupture from IVOCT: towards an AI...

The finite element method (FEM) allows for robust patient-specific simulations of atherosclerotic artery biomechanics, providing valuable insights for optimized stenting decisions. However, the complexity of the model's setup and prolonged computation times prevent FEM from being integrated into clinical practice. To enable real-time processing of image-based data into FEM-equivalent results, we developed a denoising diffusion model (DM) conditioned on labeled intravascular optical coherence (IVOCT) images, mapping the initial plaque morphology to stressed and deformed states throughout and after a percutaneous coronary intervention (PCI). By pre-training the DM on a thorough range of idealized artery cross-sections and subsequently fine-tuning feature extraction from patient-specific arteries, we ensured reliable generalization across diverse morphological variations while maintaining high-fidelity predictions. The DM achieved a structural similarity index >0.894 and learned perceptual image patch similarity <0.048, indicating high-quality image generation. Among its core functions, the DM enabled full 3D reconstruction of post-PCI morphologies, showcasing accurate representations of lumen expansion, with observed deviations of 6.86 % from FEM area profile curves. Assessing the risk of calcification and fibrous cap rupture can also be achieved by comparing stress levels with their respective failure thresholds. Consequently, clinicians can weigh options for plaque modifications and stent landing-site selection based on lesion stability under specific stent loadings. This study offers a rapid and reliable alternative to conventional FEM modeling for PCI planning, supporting its integration into routine clinical workflows.