From images to maps: a pipeline for creating synthetic electroanatomic mapping data
A Von Kietzell, V Vigneswaran, K Maciunas, N Bodagh, M Klis, A Gharaviri, A Chiribiri, N Grubb, S Haldar, M O'neill, M O Bernabeu, S E WilliamsAbstract
Background
Electroanatomic mapping (EAM) generates detailed spatial and temporal datasets describing cardiac anatomy and electrophysiological function [1]. Despite their value, EAM data are rarely stored in electronic health records or shared between centres. Available datasets are typically small, fragmented, and difficult to combine due to technical heterogeneity and privacy constraints. The lack of large, standardised, and accessible EAM data limits algorithm development, benchmarking, and reproducibility in computational electrophysiology [2]. In other domains of medical imaging, synthetic data have successfully enabled model training, validation, and benchmarking while avoiding privacy concerns [3].
Purpose
To develop and validate an open-source pipeline for generating synthetic electroanatomic mapping datasets that replicate realistic atrial geometries and local activation time maps.
Methods
Left atrial surfaces derived from manual segmentation of the blood pool from MRI data were treated as ground-truth anatomy. Synthetic EAM geometries were produced by deforming these surfaces to reproduce characteristic distortions seen in clinical maps. Deformations included large- and small-scale perturbations to mimic mapping artifacts and regional inaccuracies. Expert electrophysiologists iteratively reviewed successive synthetic versions (V1–V4), refining deformation parameters for improved realism. The final iteration incorporated region-dependent noise and a corruptive term simulating poor triangulation. Synthetic local activation time maps were generated by simulating paced activation on the original surfaces using OpenCARP [4] and EP Workbench [5], sampling LATs at 1000 virtual electrode points and interpolating these across the deformed surfaces. Realism was assessed by comparing mean curvature distributions of n=23 real and synthetic mapping meshes via the Earth Mover’s Distance (EMD) metric [6].
Results
The median EMD between real and synthetic mesh curvatures decreased across versions, indicating progressive geometric realism (Figure 1). Expert review confirmed these findings, noting improved surface fidelity and more authentic mapping artifacts. The iterative feedback process successfully integrated clinical expertise into model refinement, producing synthetic geometries closely resembling clinical mapping data.
Conclusions
We present a reproducible pipeline for generating synthetic electroanatomic mapping data combining anatomic fidelity, physiologic plausibility, and expert-informed realism. The generation of large, synthetic mapping datasets using this pipeline could support future analysis, method development and validation using EAM data at scale.Figure 1