AI-Driven Automated Processing of 4D Flow MRI Reveals Hemodynamic Changes for BAV Patients vs. Healthy Controls (RF_TH_225)

Background: Patients with bicuspid aortic valve disease (BAV) have elevated risk of aortic valve dysfunction and aortopathy ^1,2.  Clinical management of BAV patients includes regular assessment of aortic hemodynamics, including by 4D flow MRI, which can be used to measure parameters such as aortic peak velocities (V_max), pulse wave velocity (PWV), kinetic energy (KE) and energy loss (EL) ^2–6.  Traditional work-flows for performing such hemodynamic assessments require time-consuming manual intervention, which can also result in appreciable inter-observer variability affecting the measurements ⁷.  Recent work has demonstrated full automation of all 4D flow processing steps required for such hemodynamic assessment, using AI-driven techniques to replace the traditional manual interventions ⁸.  Here we investigate the impact of using such automation in a study comparing aortic hemodynamics in a large cohort of BAV patients and healthy control subjects.

Methods: Data acquired by 4D flow MRI from BAV patients (n=147) and healthy control subjects (n=101) was assembled retrospectively and processed using a fully-automated AI-driven processing pipeline (Figure 1) ⁸.  Images were acquired at multiple sites in two institutions using a 4D flow MRI sequence with these sequence parameters: 1.5 or 3 T, 1.5–4.5 mm³, 30–42 ms, R=2–8, VENC 80–300 cm/s, sagittal-oblique thoracic aorta FOV.  All subject data were accessed with IRB oversight and approval.  Processing was performed twice: once manually using traditional methods, and once automatically using AI networks trained for the manual tasks ^9,10.  Quantifications of Vmax, PWV, KE and EL for all subjects were performed using traditional algorithmic analysis ^3,5,6 after preprocessing and segmentation.

Results: BAV patients and controls were matched in age (p = 0.56), but males were more prevalent for BAV patients than for healthy controls (p < 0.005) (Table 1).  In comparison of aortic hemodynamics, V_max, KE and EL were elevated for BAV patients compared to controls, but PWV was not.  From manual data processing, V_max, KE, and EL were higher in BAV patients (respectively 1.77 m/s vs. 1.29 m/s, 5.0 μJ vs. 3.1 μJ, 0.048 μJ vs. 0.035 μJ; p< 0.005 for all) (Figure 2).  From automated data processing, the same measures were elevated (1.74 m/s vs. 1.31 m/s, 4.9 μJ vs. 3.1 μJ, 0.048 μJ vs. 0.035 μJ; p < 0.005 for all) (Figure 2).  In both manual and automated PWV quantifications, no significant differences between BAV and controls were observed (manual 6.5 m/s vs. 6.4 m/s; automated 6.6 m/s vs. 6.4 m/s).  For all parameters, when comparing automated vs. manual quantification, no significant differences were found (Figure 2).

Conclusion: The overall hemodynamic profile for these BAV patients, with elevated aortic blood velocities and energetics but unchanged pulse wave velocity, is consistent with prior studies of BAV hemodynamics.  Differences of hemodynamics observed for BAV patients vs. healthy control subjects were preserved by analysis using fully-automated processing of 4D flow MR images, including elevation of peak velocities and energy metrics for BAV patients but generally unchanged pulse wave velocities.  These hemodynamic changes are indicative of the valve-mediated hemodynamic disruptions known to occur in BAV.  The automation of all processing steps allows for more reproducible quantifications, and it enables analysis of very large cohorts, for which the effort required to perform traditional, manual data processing would otherwise be prohibitive.