Abstract
INTRODUCTION: Cardiac fibrosis influences atrial fibrillation (AF) progression and ablation outcomes, with late gadolinium enhancement (LGE) MRI providing a non-invasive tool to measure fibrosis distributions. While deep learning (DL) has shown promise in predicting ablation success, training such pipelines is limited by the availability of real patient data. METHODS: In this study, we generated synthetic fibrosis distributions using a denoising diffusion probabilistic model trained on a collection of 100 real LGE-MRI distributions. We incorporated them into 1,000 bi-atrial meshes derived from a statistical shape model and simulated AF episodes on them before and after various ablation strategies to expand the training dataset for DL-based outcome prediction. Our approach aims to improve the predictive performance of the DL pipeline by enhancing dataset diversity and better-capturing patient variability. RESULTS: We showed that the fibrosis distributions generated by the diffusion model closely resemble real LGE-MRI distributions, based on metrics such as mean intensities ( 1.1 ± 0.2 vs. 1.1 ± 0.3 ) and average Shannon entropy ( 0.77 ± 0.06 and 0.81 ± 0.03 ). AF biophysical simulations can be effectively conducted on bi-atrial meshes incorporating these synthetic distributions. Training the deep learning pipeline on these simulations produces performance metrics comparable to those achieved with real LGE-MRI distributions (ROC-AUC = 0.952 vs. 0.943). CONCLUSION: We have shown the ability of synthetic fibrosis distributions to be a data augmentation tool for deep learning classification of outcomes of various ablation strategies, which may enable rapid and precise assessment of atrial fibrillation treatment strategies.