Segmentation of 4D Flow MRI: Comparison between 3D Deep Learning and Velocity-Based Level Sets.

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Biomedical Engineering

Background:

• Thoracic aortic aneurysms (TAAs) are life-threatening dilations of the aorta, with surgical decisions often relying on maximum diameter, a metric known for its limitations.
• Four-dimensional (4D) flow magnetic resonance imaging (MRI) enables novel biomarker calculations, such as wall shear stress (WSS), crucial for TAA assessment.
• Accurate segmentation of the aorta across the cardiac cycle is essential for reliable WSS calculation, posing a significant technical challenge.

Purpose of the Study:

• To compare the performance of two automated thoracic aorta segmentation methods using 4D flow MRI data.
• To evaluate a level set-based method utilizing velocity fields and a U-Net deep learning approach applied to magnitude images.
• To assess the impact of segmentation accuracy on wall shear stress (WSS) biomarker calculations in the systolic phase.

Main Methods:

• Two automated segmentation techniques were compared: a level set framework and a U-Net-like convolutional neural network.
• The methods were applied to 4D flow MRI data from 36 patients, focusing on the systolic phase with available ground truth.
• Segmentation accuracy was quantified using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD); WSS was also calculated and compared.

Main Results:

• The U-Net approach demonstrated statistically superior segmentation performance, achieving a higher DSC (0.92 ± 0.02 vs. 0.86 ± 0.5) and lower HD (21.49 ± 24.8 mm vs. 35.79 ± 31.33 mm) for the whole aorta.
• While the level set method showed a slightly smaller absolute difference in maximum WSS compared to ground truth, this difference was not statistically significant.
• The U-Net method's improved segmentation accuracy suggests its potential for more reliable WSS assessment in TAA patients.

Conclusions:

• Deep learning-based segmentation, specifically the U-Net approach, offers significant advantages for accurate thoracic aorta segmentation in 4D flow MRI.
• Accurate segmentation across all cardiac time steps is crucial for reliable evaluation of hemodynamic biomarkers like WSS in aortic diseases.
• The findings support the consideration of deep learning methods for comprehensive analysis of 4D flow MRI data in TAA management.