Domain Adaptation for Automatic Aorta Segmentation of 4D Flow Magnetic Resonance Imaging Data from Multiple Vendor Scanners

Abstract

The lack of standardized pipelines for image processing has prevented the application of deep learning (DL) techniques for the segmentation of the aorta in phase-contrast enhanced magnetic resonance angiography (PC-MRA). Furthermore, large, well-curated and annotated datasets, which are needed to create DL-based models able to generalize, are rare. We present the adaptation of the popular nnU-net DL framework to automatically segment the aorta in 4D flow MRI-derived angiograms. The resulting segmentations in a large database ($> 300$ cases) with normal cases and examples of different pathologies of the aorta provided from a single centre were excellent after post-processing (Dice score of 0.944). Subsequently, we explored the generalisation of the trained network in a small dataset of images (around 20 cases) acquired in a different hospital with another scanner. Without domain adaptation, only with a model trained with the large dataset, the obtained results were substantially worst than with adding a few cases of the small dataset (Dice scores of 0.61 vs 0.86, respectively). The obtained results created good quality segmentations of the aorta in 4D flow MRI, which can later be post-processed to assess blood flow patterns, similarly than with manual annotations. However, advanced domain adaptation schemes are very important in 4D flow MRI due to the large differences in image characteristics between different vendor scanners available in multiple centers.