Supervised learning-based approaches are increasingly used for vessel segmentation in 4D flow MRI. However, their widespread adoption is challenged by the need for diverse, reliably annotated datasets, sensitivity to acquisition and reconstruction settings, and the lack of fully defined ground truth. In this study, we investigated the use of realistic, fluid mechanics-informed synthetic 4D flow MRI data to train convolutional neural networks for time-resolved aortic segmentation. The availability of fully defined ground truth vessel geometries in synthetic data enabled quantitative evaluation prior to evaluation on in vivo data acquired across three scanners with varying field strengths and protocols. Four training strategies were evaluated: one using in vivo data only (R-28), and three based on increasing amounts and diversity of synthetic data (S-28, S-multi and S-full). Performance was assessed using voxel- and surface-based metrics, including Dice score (DSC), Hausdorff distance (HD), and Bland-Altman analysis. The S-full model achieved the best performance on the in vivo test dataset, with a Dice score of 0.956 ± 0.017 and a Hausdorff distance of 1.708 ± 1.473 mm, relative to reference annotations. Bland-Altman analysis of cross-sectional areas showed small biases and narrow limits of agreement, with 1.1% -10.9, 13.1 % for the ascending aorta and -1.0% -13.1, 11.1 % for the descending aorta. For the in vivo trained model (R-28) evaluated on synthetic data, relative cross-sectional area measurements yielded biases of 3.1% -10.9, 17.0 % in the ascending aorta and 4.0% -8.2, 16.2 % in the descending aorta, relative to synthetic ground truth. This work demonstrates that purely synthetic 4D flow MRI can be used to train neural networks for time-resolved aortic segmentation of in vivo 4D flow MRI data, enabling fully automatic inference and quantitative evaluation against fully defined ground truth.
Wolkerstorfer et al. (Wed,) studied this question.