Atrial Fibrosis Distribution Generation Based on the Diffusion Models

Deep learning (DL) models have the potential to accurately predict atrial fibrillation (AF) ablation outcomes based on patient-specific anatomical and physiological features, including fibrotic remodelling. However, it is challenging to collect enough data for training deep learning models using clinical data alone. With this motivation, we developed a method to artificially generate additional datasets. In this study, we aim to generate artificial atrial fibrosis distributions via diffusion models to increase our training dataset size by imitating independent personalized AF episodes. We validated the proposed method by applying a DL binary classifier, which predicts whether AF is sustained post-PVI ablation. Fibrosis and dominant frequency maps extracted from pre-ablation AF simulations were used as inputs to predict AF sustainability after PVI. We compared training a binary classifier on generated fibrosis data to training it on 100 real fibrosis distributions on the same LA anatomy. For the baseline classifier trained and tested on real fibrosis data, the ROC-AUC score was 0.96. In contrast, training the classifier using generated cases, and testing on the real fibrosis data, resulted in an ROC-AUC score of 0.92. The results indicate that the artificial fibrosis distributions correspond well with the real ones and can be used for dataset expansion. Code will be available at https://github.com/pcmlab/cinc23_qmul.