Automated Segmentation Reconstruction and Analysis of Cerebral Vasculature Flow Using 4D Flow MRI (RF_FR_303)

4D Flow MRI enables time-resolved three-dimensional blood flow imaging in the cerebral vasculature. However, its accuracy can be compromised by low spatiotemporal resolution, noise, phase wrapping, intravoxel phase dispersion, and partial volume effects. These issues vary across MRI systems and compromise data consistency, leading to misinterpretations unless corrected.

Methods:

We developed a 4D Flow MRI processing workflow (Figure 1) that automatically segments vasculature, reconstructs velocity fields, and evaluates flow-related parameters.

A new segmentation method termed the Combined Likelihood Iterative Segmentation (CLIS) method is introduced, leveraging signal magnitude and phase from the Standardized Difference of Means Velocity (SDM) to generate background likelihood probabilities [1]. These probabilities assess the likelihood of a voxel being part of the background. Segmentation is initialized using Pseudo Complex Difference (PCD) and the two probability fields are combined as a weighted sum using an optimization approach to maximize boundary smoothness [2]. The combined likelihood is dynamically thresholded, refining segmentation iteratively until convergence at each cardiac phase. The final mask includes voxels classified as flow over the majority of cardiac phases.

Velocity reconstruction utilizes divergence-free conditioning, Universal Outlier Detection, and Proper Orthogonal Decomposition [3][4][5]. We iterate using each method twice to effectively reduce noise and correct discontinuities. Reconstructed velocities are used to compute hydrodynamic parameters such as pressure, wall shear stress, and coherent structures, using an uncertainty-weighted optimization method to enhance accuracy [6][7].

Results:

CLIS was compared against PCD and SDM methods using an in vitro scan of an internal carotid artery (ICA) aneurysm. CLIS outperformed the alternatives (Figure 2), with a median segmentation error half that of PCD, as shown by sub-voxel separation from the true segmentation.

Velocity reconstruction was evaluated using synthetic 4D Flow MRI data with a velocity-to-noise ratio of 1 and velocity encoding at half the maximum CFD velocity. These settings were specifically chosen to simulate an extreme scenario, resulting in a highly noisy and phase-wrapped velocity field. Our workflow reduced errors by a factor of three and improved correlation with true velocities from 0.26 to 0.7, significantly minimizing noise and aliasing.

Conclusion:

Our automated 4D Flow MRI processing workflow delivers superior segmentation and flow reconstruction compared to existing methods. In an in vitro ICA flow test, the workflow reduced bias by 50%. Additionally, velocity reconstruction yielded substantial error reduction and improved correlation with true velocities. We plan to further validate our methodology using in vivo 4D Flow datasets from a variety of MRI scanners and perform comparisons with high-resolution in vitro Particle Tracking Velocity.