Joint magnitude and phase neural network segmentation for cine 2D Flow MRI

Motivation: Neural network-based segmentations of cine 2D Flow MRI can be compromised by cardiac-phase dependent inflow contrast variations and artifacts in magnitude data. Goal(s): To evaluate if neural network-based segmentation is improved by incorporating both magnitude and phase data. Approach: Masks of the ascending and descending aorta were manually generated from 213 cine 2D Flow MRI datasets. Separate convolutional neural networks (nnUNet) were trained using either magnitude data only or combined magnitude and phase data. Segmentation performance was assessed relative to manual segmentation masks. Results: State-of-the-art cine 2D Flow MRI segmentation was achieved. The addition of phase data did not significantly benefit segmentation performance. Impact: This work demonstrates state-of-the-art neural network-based segmentation of aortic vessels across the cardiac cycle using the nnUNet. The addition of phase data appears non-essential. The trained models will be made available, to promote reproducibility and accessibility in the field.