The Effect of Magnetic Resonance Image Resolution and Amount of Segmentation Smoothing on Aortic Computational Fluid Dynamics (RF_FR_343)

Magnetic resonance imaging (MRI) can be used to create patient-specific models. Typically, these are made from high resolution scans. These models could potentially be used to make a diagnosis or determine if a procedure might be beneficial. We hypothesized that image resolution the amount of smoothing done on a segmentation impacts the accuracy of aortic computational fluid dynamics (CFD) models.

We analyzed the effect of MRI resolution and segmentation smoothing on patient-specific aortic hemodynamics, specifically velocity and wall shear stress (WSS) profiles, obtained through CFD simulations. This was done to determine what values for both resolution and smoothing can create an accurate model. This analysis highlights the pivotal role of MRI resolution and segmentation quality in the effectiveness of the patient-specific model.



Methods:

An automated, isotropic, 4D whole thoracic protocol (AutoCMR: free-running, GRE, TR/TE = 4.0/2.9 ms, flip angle = 12°, resolution = 1.6mm isotropic, matrix = 192 x 192 x 192 x 30) was used to acquire MRI data of 10 patients’ torsos (4M, 62.4 ± 9.8 yr) with a 3T MR system. The systolic aortas were manually segmented in 3D Slicer and smoothed with a 12 mm kernel.

The resolution was reduced in the short axis with Python from 1.6 mm to 3.2 mm and 8.0 mm. The slice thickness of 8.0 mm mimics the MRI standard of care. Each of these models was smoothed with kernel sizes 9 mm, 15 mm, and 21 mm (Fig 1).

The models were meshed and simulations in SimVascular were performed. Laminar blood flow was simulated through 3D incompressible Navier-Stokes equations. The walls were assumed to be rigid and non-slip. The inlet boundary conditions were a time-varying flow rate, and the outlet conditions were governed by a pressure wave generated with svZeroDSolver.

The end of the diastolic phase was found for each simulation. The results at this point were segmented into three parts, and Wilcoxon signed-rank tests were performed to determine statistical significance from the reference model (p < 0.05).



Results:

We determined through visual comparison that there are qualitative differences caused by the variations in image resolution and smoothing kernel size for both the velocity and WSS (Fig 2).

We compared the velocity and WSS of the models. Wilcoxon signed-rank tests were performed comparing the models on resolution and smoothing kernel (Tbl 1).



Conclusion:

The presence of differences in WSS depended on the section, resolution, and smoothing kernel. This supported the hypothesis that kernel size affects model accuracy. Differences occurred mostly in the aortic arch.

There was no significance in all but one of the velocity tests. However, a visual inspection showed that the streamlines of the models depended on resolution and smoothing kernel.

One limitation of this study is the relatively small sample size. However, the results we have indicate that both image resolution and smoothing kernel size have an impact on the accuracy of aortic CFD models.