A Diffusion Model for Simulation Ready Coronary Anatomy with Morpho-skeletal Control

TL;DR

The paper introduces a diffusion-based framework for generating and editing coronary artery anatomies with mid-level morpho-skeletal constraints, enabling physics-based counterfactual simulations of device deployment. It advances Latent Diffusion Models by incorporating a topological regularization loss, explicit morphological and skeletal conditioning, and novel adaptive null guidance to improve conditioning fidelity while maintaining efficiency. The approach demonstrates improved topological, morphological, and skeletal fidelity on 3D multi-material coronary segmentations and enables disentangled control and editing for virtual angioplasty studies. This work provides a flexible, controllable alternative to purely parametric or image-derived digital twins, facilitating mechanistic insights into how anatomical variation impacts interventional outcomes. The framework stands to accelerate device design, trial planning, and personalized simulation in interventional cardiology by enabling targeted, topology-aware anatomy generation and editing.

Abstract

Virtual interventions enable the physics-based simulation of device deployment within coronary arteries. This framework allows for counterfactual reasoning by deploying the same device in different arterial anatomies. However, current methods to create such counterfactual arteries face a trade-off between controllability and realism. In this study, we investigate how Latent Diffusion Models (LDMs) can custom synthesize coronary anatomy for virtual intervention studies based on mid-level anatomic constraints such as topological validity, local morphological shape, and global skeletal structure. We also extend diffusion model guidance strategies to the context of morpho-skeletal conditioning and propose a novel guidance method for continuous attributes that adaptively updates the negative guiding condition throughout sampling. Our framework enables the generation and editing of coronary anatomy in a controllable manner, allowing device designers to derive mechanistic insights regarding anatomic variation and simulated device deployment.

Figures (18)

• Figure 1: We propose to control the generation of 3D multi-component diseased coronary arteries with mid-level representations such as cross-sectional morphology and tree-like skeletal structure. Synthetic arteries enable physics-based counterfactual reasoning through comparative virtual intervention studies. We synthesize semantic segmentation maps composed of luminal (blue), arterial wall (yellow) and calcified (green) tissues.
• Figure 2: We regularize the latent space $\mathbf{z}$ with a topological interaction loss applied to the calcium and lumen segmentations.
• Figure 3: Left: We condition the training process of a latent diffusion model through channel-wise concatenation of the morphological ($\mathbf{y}_m$) and skeletal ($\mathbf{y}_s$) maps to the noised latent representation ($\mathbf{z}_\sigma$). Right: Our proposed guidance algorithm operates by specifying an null skeletal condition ($\mathbf{y}^\varnothing_s$) and adaptively updating a null morphological condition ($\mathbf{y}^\varnothing_m$) based on the error between the target condition ($\mathbf{y}_m$) and the morphology exhibited by the current denoiser output ($\mathbf{\hat{y}}_m$).
• Figure 4: Longitudinal and cross sectional slices of coronary segmentation maps. First column shows that patient-specific segmentation maps from the training set exhibit several topological defects such as lumen and calcium tissues not being fully contained within vessel wall tissue (white arrows). Second and third columns show the effect of autoencoding coronary segmentation maps with (right) or without (center) using a topological loss $\mathcal{L}_T$ during training.
• Figure 5: Example morphological features, skeleton depth maps and synthetic segmentation map cross sections for various guidance methods. Filled in regions within morphological plots indicate standard deviation over 10 generated segmentation maps. A guidance weight of 5 is used when applicable. Our proposed guidance method (ANG) improves conditional fidelity while maintaining good visual quality.