Abstract
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.
Introduction
Atrial fibrillation (AF) remains one of the most prevalent cardiac arrhythmias, characterized by rapid and irregular electrical activity in the atria. Catheter ablation, a common therapeutic intervention, aims to disrupt these irregular signals by creating targeted scar tissue. However, clinical success rates remain suboptimal, ranging between 50% and 60%. This variability is largely attributed to the unique patient-specific atrial anatomy and the complex distribution of fibrosis (scarring), which significantly influences electrical propagation.
Deep learning (DL) offers a promising avenue for predicting patient-specific ablation outcomes. However, the development of robust DL models is hindered by the scarcity of high-quality Late Gadolinium Enhancement (LGE) MRI data and the computational intensity of biophysical simulations required to generate "ground truth" labels. This study investigates the use of Denoising Diffusion Probabilistic Models (DDPMs) to generate synthetic fibrosis distributions as a data augmentation strategy to overcome these bottlenecks.
Methods
Generative Modeling of Fibrosis
The researchers employed a DDPM, a generative architecture similar to those used in state-of-the-art image synthesis, to learn the underlying distribution of atrial fibrosis from 100 real LGE-MRI scans. The model was trained to reverse a diffusion process, effectively learning to generate high-fidelity synthetic fibrosis maps from random Gaussian noise.
Virtual Patient Cohort and Biophysical Simulation
To create a diverse "virtual patient" cohort, the generated 2D fibrosis distributions were projected onto 1,000 unique 3D bi-atrial meshes derived from a statistical shape model. Electrophysiological activity was then simulated using the openCARP framework. For each virtual patient, four distinct ablation strategies—Pulmonary Vein Isolation (PVI), PVI plus Left Atrial (LA) ablation, PVI plus Right Atrial (RA) ablation, and a comprehensive PVI+RA+LA approach—were tested to determine their efficacy in terminating AF.
Predictive Pipeline and Multi-modal Fusion
A Siamese deep learning architecture was developed to predict ablation outcomes. The model processed 2D feature maps representing fibrosis (F), Phase Singularity Density (PSD), and Dominant Frequency (DF). To effectively integrate information from both the left and right atria, the study introduced advanced fusion modules, including the Multi-modal Outer Arithmetic Block (MOAB) and the Flattened Outer Arithmetic Attention (FOAA) block.
Results
Fidelity of Synthetic Data
The synthetic fibrosis distributions demonstrated high statistical similarity to real clinical data. Mean intensity values were nearly identical ($1.1 \pm 0.2$ IIR for synthetic vs. $1.1 \pm 0.3$ IIR for real). Furthermore, Shannon entropy analysis confirmed that the synthetic images captured the structural complexity of real-world fibrosis, with an average entropy of $0.77 \pm 0.06$ compared to $0.81 \pm 0.03$ for real scans.
Predictive Performance
The DL model trained exclusively on synthetic data achieved an ROC-AUC of 0.943, performing remarkably close to the model trained on real clinical data (ROC-AUC 0.952). In contrast, models trained on random Gaussian noise distributions performed significantly worse (ROC-AUC ~0.68), highlighting that the specific spatial patterns captured by the DDPM are essential for accurate electrophysiological prediction.
Discussion
A key finding of this study is the "physics bridge"—the observation that synthetic data generated by a vision-based model is sufficiently realistic to drive complex biophysical simulations. By using Shannon entropy as a quality filter, the researchers were able to ensure that only high-fidelity synthetic distributions were used for training, effectively mitigating the risk of model "hallucinations."
While the results are promising, the study acknowledges limitations, including the reliance on a relatively small number of base healthy anatomies for mesh generation and the use of a standardized AF initiation protocol. Future work will focus on expanding anatomical diversity and validating the pipeline against prospective clinical data.
Conclusion
This study demonstrates that synthetic fibrosis distributions generated via DDPMs can serve as a powerful data augmentation tool. By bridging the gap between generative vision models and biophysical simulations, this approach enables the training of high-performance predictive models for AF ablation outcomes, potentially paving the way for more personalized and effective treatment strategies.
Summary updated