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

Jonas Leite1, Louis Parker2, Marie Shannon Soulez3, Tom Da Silva-Faria4, Khaoula BOUAZIZI5, Perrine Marsac3, Elie Mousseaux6, Moussa Gueda Moussa4, nicolas badenco7, estelle gandjbakhch8, alban redheuil8, Mikaël Laredo9, Emilie Bollache10, Nadjia Kachenoura11
1Sorbonne Universiter, 2Laboratoire d'Imagerie Biomédicale (LIB), Sorbonne Université, Institut National de la Re-cherche Médicale (INSERM), Centre National de la Recherche Scientifique (CNRS), Paris, France., 3Sorbonne Université, 4sorbonne universite, 5LIB, 6European Hospital Georges Pompidou, Assistance Publique - Hôpitaux de Paris (AP-HP)/PARCC,Université Paris-Cité, INSERM, 7Institute of Cardiometabolism and Nutrition (ICAN)/Institut de Cardiologie, Pitié-Salpêtrière Hospital, 8sorbonne universite/Institute of Cardiometabolism /Unité d'Imagerie Cardiovasculaire et Thoracique (ICT), Pitié-Salpêtrière Hospital and Nutrition (ICAN)-, 9Sorbonne Université/ Institute of Cardiometabolism and Nutrition (ICAN)/Unité d'Imagerie Cardiovasculaire et Thoracique (ICT), Pitié-Salpêtrière Hospital, 10Inserm, Laboratory of Biomedical Imaging, 11INSERM

Abstract

While previous 4D flow MRI segmentation methods mostly focused on the aorta, only few were proposed for the left atrium (LA) whose geometry and flow are more complex. We aimed to design a deep learning (DL) pipe-line for automated LA segmentation from 4D flow MRI, which is crucial to access function and hemodynamics especially in atrial fibrillation (AF). We included 59 patients (3D LA annotations on all frames through the full cardiac cycle on n=16) with AF scanned on a 1.5T Siemens MRI, and 28 healthy controls (n=9 full annotations) acquired at 3T (General Electric). We derived unannotated cases from corresponding CT segmented LA masks which were available in AF patients at systole and/or diastole, using spatial registration onto the 4D flow MRI volume and then throughout the cardiac cycle using Elastix algorithm. With such approach we generated ground truth 3D+time LA annotations. To reduce registration errors, our proposed DL model which included two 3D ResUnet architectures was trained on restricted time-frame window around the reference (initial±3 frames; total N=638 training volumes from 43 AF patients and 18 controls using data augmenta-tion). On the testing set (16 patients and 9 controls), our DL model achieved good accuracy at the initial time-frame (Dice similarity coefficient DSC=0.85±0.06/Hausdorff distance HD=4.15±1.41 mm), with better per-formance in patients (DSC=0.87±0.03/HD=3.73±0.89 mm) than in controls (DSC=0.82±0.07/HD=4.80±1.76 mm), while across all time frames, perfor-mance declined (averaged DSC=0.81±0.07/HD=8.50±4.94 mm), likely due to limited 3T data and external time frames during training. Finally, predicted LA volumes (r=0.85/relative bias µ=6.06±8.67%) and hemodynamic param-eters (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) were well correlated with derived expert values. These results suggest promising results for the automated, accurate LA segmentation and biomarker quantification from 4D flow MRI using our DL pipeline.