Computerized Analysis of the Human Heart to Guide Targeted Treatment of Atrial Fibrillation.

Current treatment for atrial fibrillation (AF) remains suboptimal due to a lack of understanding of the atrial substrate which sustains AF in human atria. Comparing atrial structural characteristics (fibrosis, myofibers and wall thickness) across different hearts is challenging due to the complexity and thinness of the atrial wall. There is also a need for quantitative tools to guide treatment strategies in clinical settings. Using 111 late gadolinium-enhanced MRI (LGE-MRI) scans taken from patients with AF from two clinical centers, we developed novel convolutional neural networks to perform automatic bi-atrial segmentation. Then a standardized 2D representation of atrial anatomical structures was used for analyzing and comparing patient data. We also developed an algorithm to define rulebased fibers in volumetric human atrial models using existing knowledge about atrial fibers. Finally, realistic fibrosis and the generated fibers were integrated into voxelized atrial geometries sourced from the LGE-MRIs. Computer simulations were performed on these geometries, using techniques that made efficient use of computational memory. The techniques also allowed for rapid and reliable reentry formation, reducing computational time. Our analysis of three flattened atria concisely showed that the lateral right atrium and the inferior pulmonary veins accrued substantial fibrosis. The fiber generation method gave smooth and finely tunable results. The simulations showed that fibrosis and fibers substantially affect AF dynamics. The proposed pipeline can potentially identify biomarkers and conduct computer simulations of the human heart to guide targeted ablation treatment.