Coupling synthetic and real-world data for a deep learning-based segmentation process of 4D flow MRI

Area of Science:

• Medical Imaging
• Computational Fluid Dynamics
• Artificial Intelligence

Background:

• 4D flow MRI provides vital blood velocity and morphological data, primarily for large vessels like the thoracic aorta.
• Segmentation of 4D flow MRI data is challenging and time-consuming.
• Limited availability of 4D flow MRI datasets hinders neural network performance due to the technique's recent clinical adoption.

Purpose of the Study:

• To develop a pipeline for generating synthetic thoracic aorta PC-MRA data.
• To augment limited real-world PC-MRA datasets.
• To enhance the accuracy of neural network-based segmentation of the thoracic aorta.

Main Methods:

• A statistical shape model synthesized artificial geometries to increase data variability.
• Computational fluid dynamics simulations generated velocity fields.
• Synthesized volumes, combined with real data, trained a 3D U-Net neural network.

Main Results:

• The inclusion of synthetic data significantly improved segmentation performance (DICE score of 0.83) compared to using only real data (DICE score of 0.65).
• Enhanced target reconstruction was achieved with the combined dataset.

Conclusions:

• The proposed pipeline effectively increases the numerosity and variability of PC-MRA datasets.
• Synthetic data generation improves thoracic aorta segmentation accuracy using PC-MRA.