Transfer learning for automatic aorta segmentation in 4D-Flow magnetic resonance imaging data

Abstract

The lack of standardized pipelines for image processing, together with the noisy nature of data, challenges the application of deep learning (DL) techniques for the segmentation of cardiovascular structures in 4D-flow magnetic resonance imaging (MRI) data. Furthermore, DL-based algorithms require large, well-curated and annotated datasets for training, which is not straightforward. We therefore present a transfer learning approach to automatically perform aortic segmentation with 4D-flow MRI contrast-free data coming from different clinical sites and acquisition machines. Three datasets were considered: VH1, VH2 (only used for testing) and CAMRI. Two convolutional neural networks, based on the nnU-net framework, were trained with manual segmentations: LD (trained on VH1) and SD (trained on CAMRI). Performance was assessed using Dice (DS) and Jaccard score, Haussdorf distance, and average symmetrical surface distance. Transfer learning was applied using LD network to predict CAMRI data. LD network segmentations of VH2 and CAMRI datasets showed a median DS of 0.944 and 0.700 respectively. TL, using only a small fraction of CAMRI data for training, improved LD network generalization capabilities, increasing DS to 0.868. TL performance was comparable to SD network trained on all CAMRI data (0.897 DS). We demonstrate the applicability of the nnU-net framework for fast and automated 3D aortic segmentation in 4D-flow MRI datasets from different clinical sites with a comparable state-of-the-art performance, even without the need of contrast. TL approach increased generalisation, thus suggesting that it can reduce the tedious and time-consuming human intervention in the segmentation process, facilitating the availability of large annotated databases.