Recent literature has proposed various methods for automatic segmentation of the thoracic aorta in 4D flow MRI at the systolic peak, often evaluated on homogeneous or single-center cohorts. In this study, we compared the performance of three state-of-the-art deep learning models, including the nn-UNet, on heterogeneous 4D flow MRI data acquired using two MRI scanners, two field strengths (1.5T and 3T), across two centers, including both healthy volunteers and patients with aortic diseases. The dataset comprised 143 individuals (55 women, age 57 ± 16 years: healthy = 55, type 2 diabetes = 30, myocardial infarction = 33, ascending aorta aneurysm = 25). Reference annotations (GT) were created in 3D using validated software (Mimosa, Sorbonne Université) on systolic peak angiograms from 4D flow MRI. The dataset was split into training (n = 100) and evaluation subsets (n = 43). The models were assessed based on Dice similarity coefficient (DSC), average symmetric surface distance (ASSD), and ascending (AAo) and descending (DAo) aorta mean and maximum diameters. The nn-UNet achieved the highest performances (DSC = 0.88 ± 0.05, ASSD = 1.52 ± 0.96 mm, AAo and DAo mean and max diameters Pearson correlation with GT r ≥ 0.97). While the nn-UNet demonstrated superior accuracy, all three models delivered competitive segmentation results, suitable for aortic morphology assessment.