Estimating heterogeneous conduction velocities (CVs) in the atria is essential for understanding arrhythmia mechanisms but remains challenging due to sparse and noisy clinical data. FiberNet, a physics-informed neural network-based method, offers a data-efficient approach to estimate direction-dependent CVs and fiber orientation from local activation time maps.
We present a synthetic 2D benchmark setup, physiologically motivated by atrial tissue properties, specifically the preferential orientation of cardiomyocytes, as well as region-specific CVs and anisotropy ratios. By systematically varying data quality and fiber complexity, we analyzed FiberNet's robustness and accuracy in estimating anisotropic properties across heterogeneous tissue, accounting for anatomical variability.
For uniform or sharply heterogeneous tissues with sparse data, 84%, 60%, and 74% of the predictions were within error thresholds of 30° for fiber angle α, 0.1 m/s, and 0.2 m/s for CV, respectively. Prediction accuracy decreased under Gaussian noise on activation times (66%, 18%, and 36% within the same error bounds for standard deviation σ = 1 ms). In regions with gradual fiber transitions, only 57% of α predictions remained below 30°. These results highlight conditions of reliable performance and opportunities for targeted improvement.