Neural Networks for Left Atria Shape Completion from Sparse Catheter Paths

Background: Cardiac arrhythmia is the clinical term for the set of diseases wherein the heart beats irregularly. Of these conditions, atrial fibrillation is one of the most prevalent, increases the risk of stroke fivefold, impacts quality of life, causes hundreds of thousands hospitalizations in the US and is linked with increased mortality. Electrical pulmonary vein isolation from the left atria (LA) body is performed using ablation for treating AF. The procedure will hasten significantly if it was guided by an LA shape visualization early on.

Methods: We propose a method to reconstruct the shape of the left atria during the electrophysiology procedure from a series of simple catheter maneuvers. Synthetic LA sampled from a statistical based model and generated catheter paths, are used to train a denoising auto encoder to reconstruct the LA shape. Those paths mimic realistic ones which were later performed in a lab phantom. A novel spatial weight smoothing term is introduced to learn smooth shape parts. We compare the results against training from partial data generated by the intersection of a randomly generated sphere and the atria.

Results: The network successfully reconstructs the LA shape solely based on the given paths achieving around 0.85,0.75 DICE score for test set and lab phantoms, respectively. The augmented loss term shows qualitative and quantitative improvement in the results.

Conclusions: LA shapes were obtained at a minute time instead of 10-15 minutes, required for current clinical practice. This could be beneficial to many cardiac minimally invasive treatments.