Abstract 15172: Improving Cardiac Segmentation for Atrial Fibrillation Ablation: A Prospective Trial of Machine Learned Geometric Dissection vs Experts

Abstract

Introduction: Segmenting cardiac computed tomography (CT) to provide anatomic guidance for Atrial Fibrillation (AF) ablation is routinely applied, but is time-consuming and prone to error. Machine learning (ML) is a powerful approach that could automate this approach, but is hindered by the small size of available labeled datasets. Hypothesis: We hypothesized that a new computational pipeline, in which an ML model is trained mathematically in a small cohort (N=20) using geometrical heart avatars derived from computer graphics imaging (CGI), rather than on manually-segmented data, would enable rapid expert-level segmentation of raw cardiac CT scans. Methods: We first encoded anatomical knowledge with generic geometrical avatars and derived a “virtual dissection” method to geometrically parse the heart (Fig A). An ML model trained by virtual dissection using 20 cases was able to rapidly and accurately segment the pulmonary veins (PVs), left atrial appendage (LAA), and left atrium (LA) from cardiac CT scans (Fig B), which we tested in a retrospective study of N=100 patients (30% women, 64.7±10.1Y) and in a prospective clinical trial of N=42 patients (42.9% women, 65.2±10.8Y) undergoing AF ablation, against a panel of 3 experts (Fig C). Results: In a retrospective study (N=100), ML achieved median Dice scores of 96.6% (IQR: 95.1% to 97.5%), similar to experts (p<0.05). In a prospective study (N=42), this pipeline reduced segmentation time (2.3±0.8 vs 15.0±6.9 minutes; p < 0.00001; Fig D), but achieved similar Dice scores (93.9% (IQR: 93.0% to 94.6%) vs 94.4% (IQR: 92.8% to 95.7%); p<0.05; Fig E). Conclusions: In our prospective trial, virtual dissection (machine learning of cardiac structures based on mathematical cardiac geometry) accelerated cardiac segmentation prior to AF ablation. In general, this approach may reduce the dependence on large training datasets for machine learning, and could be applied to other organ systems for diverse therapeutic strategies.