Population-based personalization of a 2D diffusion-based model of myocardial infarct
Résumé
Simple and effective infarct models are needed to enhance cardiac biophysical models or generate more accurate ground truth data for machine learning applications. This study presents a diffusion-based approach for modeling infarct shapes, applied to generate synthetic 2D infarct images. The diffusion is non-uniform, with different diffusion coefficients assigned to various sectors of the myocardium. We integrate a population-based approach to learn the model parameters and personalize them to a real population. This personalization method is relevant for models with randomness, such as our model, and is achieved through a gradient-free algorithm (CMA-ES), optimizing a distribution similarity metric between real and synthetic populations. We assess this approach on 2D infarct segmentations from LGE MR images and aligned to a reference geometry, from 117 patients with acute myocardial infarct. We generate 500 synthetic images and evaluate their characteristics using an attribute-based variational auto-encoder (AR-VAE), which captures a latent space with key infarct characteristics specifically disentangled on specific dimensions (transmurality, size, and orientation). Our model offers a simple, physically-based method to generate infarcts with diverse shapes. The evaluation shows that the synthetic population accurately reflects the distributions of the key infarct characteristics, with more sophisticated infarct shapes compared to previous simpler models also personalized with the population-based approach.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |