Deep Learning-Based Dynamic Segmentation of the Left Atrium in 4D Flow MRI.
Authors
Affiliations (7)
Affiliations (7)
- Laboratoire d'Imagerie Biomédicale (LIB), Sorbonne Université, INSERM, CNRS, Paris, France.
- Institute of Cardiometabolism and Nutrition (ICAN), Paris, France.
- European Hospital Georges Pompidou, Assistance Publique - Hôpitaux de Paris (AP-HP)/PARCC, Université Paris-Cité, INSERM, Paris, France.
- Centre Cardiologique du Nord Saint-Denis, Saint-Denis, France.
- Research & Clinical Translation, Magnetic Resonance, Siemens Healthineers AG, Erlangen, Germany.
- Sorbonne Université, AP-HP, Institut de Cardiologie, Hôpital Pitié-Salpêtrière, Paris, France.
- Sorbonne Université, AP-HP, Imagerie Cardiovasculaire et Thoracique, Institut de Cardiologie, Hôpital Pitié-Salpêtrière, Paris, France.
Abstract
To design a deep learning pipeline for automated, time-resolved segmentation of the left atrium (LA) from 4D flow MRI data. We studied 100 individuals including 65 patients with atrial fibrillation (AF) and 35 healthy subjects (HS) resulting in 2530 4D flow MRI time-volumes acquired in different centers and scanners. In AF patients, expert data annotation was initialized on a single frame of highly resolved computed tomography data and then registered (initialization frame), propagated through time and corrected manually on 4D flow MRI, while it was performed manually in HS. Segmentation pipeline was based on a two-stage nnU-Net and was evaluated through segmentation performance metrics, as well as concordance of LA volumes through time and inner LA hemodynamics between prediction and reference segmentations. Testing on 35 AF patients and 15 HS revealed high segmentation performances (overall Dice score = 0.89 ± 0.03 for the initialization frame and 0.86 ± 0.04 for all time frames; AF patient/HS- and vendor-based Dice scores ≥ 0.86 ± 0.03 for the initialization frame and ≥ 0.83 ± 0.04 for all time frames). Strong associations and low Bland-Altman biases were obtained between predicted and reference LA volumes (r ≥ 0.89), velocities (r ≥ 0.96), stasis (r ≥ 0.99), vorticity (r ≥ 0.97), and kinetic energy (r ≥ 0.94). Interestingly, associations remained strong when comparing prediction against 4D flow-independent cine SSFP-derived LA volumes (r ≥ 0.87). The proposed time-resolved nnU-Net pipeline robustly and accurately segments the LA from 4D flow MRI images demonstrating excellent segmentation performance, strong agreement with reference volumes and velocity-based indices, as well as reliable generalization across vendors and patient populations.