Synthetically trained convolutional neural network vessel segmentation for 4D Flow MRI

Motivation: Segmenting 4D Flow MRI is essential for automatic flow data post-processing. Training neural network-based segmentation is restricted by the limited availability of data and annotations. Goal(s): To investigate whether neural network segmentation can effectively be trained on synthetic 4D Flow MRI data. Approach: Synthetic 4D Flow MRI datasets, based on realistic aortic flow simulations, embedded within in-vivo backgrounds were used to train the nnUNet model using: (1) in-vivo data only, (2) synthetic data only, (3) combined in-vivo and synthetic data. Results: Aortic segmentation in 4D Flow MRI is feasible using neural networks trained on synthetic data. Impact: This work demonstrates the advantages of synthetic datasets for training neural network segmentation of aortic vessels. The generated data and trained models will be made available to promote reproducibility and segmentation accessibility in the field.