Deep Learning-Based Dynamic Segmentation of the Left Atrium in 4D Flow MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Accurate segmentation of the left atrium (LA) is crucial for assessing cardiovascular health, particularly in patients with atrial fibrillation (AF).
• Traditional manual segmentation of 4D flow MRI data is time-consuming and prone to inter-observer variability.
• Developing automated methods is essential for efficient and reproducible analysis of LA dynamics.

Purpose of the Study:

• To design and validate a deep learning pipeline for automated, time-resolved segmentation of the left atrium (LA) from 4D flow MRI.
• To evaluate the pipeline's performance in segmenting LA across different patient populations (AF patients and healthy subjects) and imaging vendors.
• To assess the concordance of derived hemodynamic parameters between the automated segmentation and reference data.

Main Methods:

• A two-stage nnU-Net deep learning architecture was employed for segmentation.
• The study included 100 individuals (65 AF patients, 35 healthy subjects) with 2530 4D flow MRI time-volumes from multiple centers and scanners.
• Segmentation performance was evaluated using Dice scores, and agreement was assessed for LA volumes, velocities, stasis, vorticity, and kinetic energy.

Main Results:

• The pipeline achieved high segmentation performance with an overall Dice score of 0.89 ± 0.03 for the initialization frame and 0.86 ± 0.04 for all time frames.
• Strong associations (r ≥ 0.89) and low biases were observed between predicted and reference LA volumes and hemodynamic indices.
• The segmentation demonstrated robust generalization across different vendors and patient groups, maintaining strong associations even when compared to 4D flow-independent cine SSFP data.

Conclusions:

• The proposed time-resolved nnU-Net pipeline offers a robust and accurate solution for automated LA segmentation from 4D flow MRI.
• The method demonstrates excellent segmentation performance and strong agreement with reference volumes and hemodynamics.
• The pipeline shows reliable generalization capabilities, making it suitable for diverse clinical applications and research settings.