• 4D flow
• aorta
• deep learning
• MRI
• neural network
• segmentation
• semantic segmentation

The main objective of this thesis, carried out at the BioCardioLab of the Foundation Gabriele Monasterio in Massa, is to realize a deep neural network that is able to fully automate the 3D segmentation of the aorta from 4D Flow PCMRA images.
The starting point of this work are the 3D volumes generated by the 4D Flow MRI acquisition and divided into 4 datasets: one of 'magnitude' and the three time-resolved velocity datasets along the three spatial directions. They are processed and used to obtain a single 3D PCMRA dataset. In the course of this thesis work, multiple network architectures, based on the classical U-net architecture, and optimization strategies have been experimented.
In particular, the performance of a 3D and a 2D network has been compared on the same test dataset.
Several network performance optimization strategies were used, such as Data Augmentation on volumetric images and the creation of two-channel PCMRA volumes that exploit the information derived from the systolic and diastolic phases of the cardiac cycle for a better visibility of cardiac structures.