Deep Learning-Based Left Atrial Segmentation and Hemodynamics Estimation using 4D Flow MRI

We designed a deep learning (DL) pipeline for automated left atrial (LA) segmentation, which is crucial to evaluate function and hemodynamics from 4D flow MRI. We included 59 patients (1.5T) with atrial fibrillation (AF) and 27 healthy controls (3T) who underwent MRI. We implemented and trained a dual 3D ResUnet architecture to crop and segment the LA. Results on the testing set (N=16 patients/9 controls) revealed good DL performances against ground truth (GT) expert masks on the time frame used for GT initialization (mean Dice score, DSC=0.85±0.06; Hausdorff distance, HD=4.15±1.41 mm), with higher performances for the more represented AF patients (DSC=0.87±0.03, HD=3.73±0.89 mm) than controls (DSC=0.82±0.07, HD=4.80±1.76 mm). A drop in DSC was observed across all time frames on which GT was registered (all data: DSC=0.81±0.07, patients: DSC=0.84±0.03, controls: DSC=0.77±0.08), mainly attributed to limited 3T data and external time frame in training. Finally, extracted LA volume and hemodynamic indices showed good agreement with expert-derived values (LA volume: r=0.85, bias µ=6.06±8.67%, main vortex eccentricity: r=0.88, µ=0.55±9.23%, vortex amplitude evaluated through λ₂ and Q criteria: r = 0.94, µ=-12.47±9.25% and r=0.91, µ=-16.27±10.95%, respectively).