Machine-Learned Hemodynamic Estimates from Contrast-Enhanced MRA and Direct Velocity Measurements from 4D flow MRI Are Equally Predictive of Adverse BAV Outcomes

Background: Bicuspid aortic valve (BAV) carries elevated risk of aortopathy/complications^1,2, but current risk assessment standards use only CT/MR diameter measurements, limiting predictive performance. Aortic hemodynamics—e.g., peak velocity (V_max) and wall shear stress (WSS)—are altered in BAV aortopathy^3,4, but quantifying them requires time-consuming 4D flow MRI (4DF) and is not routinely done⁵. Contrast-enhanced MR angiography (CEMRA) is frequently used clinically⁶ but does not assess hemodynamics. Recent machine learning (ML) work showed cycleGAN networks can estimate velocity directly from CEMRA⁷. Here, we applied these methods in a cohort of BAV patients, hypothesizing that ML-derived velocities improve prediction of adverse aortic outcome (AAO: dissection, aortic surgery) and perform similarly to direct in vivo 4DF measurement.

Methods: Imaging from BAV patients (no prior surgery) 2011–2024 was retrospectively accessed. All patients had CEMRA and 4DF. AAO was tabulated by electronic health record (EHR) query or chart review if no programmatic EHR match. In total, 160 patients had CEMRA/4DF and ≥1y follow-up. Patients were classified AAO from earliest record of adverse outcome or stable if none found (Figure 1).



A cycleGAN to estimate aortic velocities from CEMRA (fluid-physics 3D dense U-net, previously trained by n=775 BAV 4DF/CEMRA)⁷ was applied in the 160 patient CEMRA data (no overlap with training set). An automated pipeline tool^8,9 was used to process 4DF data. Ascending aorta V_max and maximal WSS were quantified from both ML-derived and 4DF-measured data (Figure 2).



CycleGAN-derived V_max or WSS, patient age, and aortic diameter (mid-ascending) were used to fit linear prediction models (binomial). For each hemodynamic parameter, two models were tested, including or excluding diameter as a variable. A diameter/age-only model was also fit. Predictive performance was evaluated by area-under-curve (AUC) with bootstrapped confidence intervals (CI). Finally, the same analysis was performed using 4DF V_max and WSS.

Results: The cohort had age 49.1y(±14.5y) and 74% male sex; follow-up was 7.2y(±3.8y); 101 patients were stable and 32 had AAO (Figure 1). Predicting AAO from CEMRA hemodynamics had high performance (AUC=0.77–0.82; Figure 3). Predicting AAO with 4DF measurements performed similarly (AUC=0.75–0.84; Figure 3) and not significantly different (AUC CI [0.64,0.85]/[0.66,0.88] 4DF vs. [0.65,0.87]/[0.71,0.84] MRA). For both imaging types, either hemodynamic predictor was better than diameter alone (AUC=0.55, CI [0.45,0.64]).

Conclusion: Diameter-only AAO prediction performed poorly, only marginally better than random chance. Introducing velocity information, whether inferred from CEMRA by cycleGAN or measured from 4DF, allowed significantly better prediction of adverse outcomes. The strong performance predicting adverse outcome using hemodynamic quantifications from widely-available CEMRA images suggests high potential utility for BAV aortopathy management and risk stratification.

Figure 1. Summary of demographics for the overall cohort and the adverse-outcome subgroups (top), and electronic health record keys used to classify outcomes (bottom).
Figure 2. Schematic for data processing in this study, with a processing pipeline for CEMRA (top) using physics-informed cycleGAN to estimate velocities, or 4D flow MRI (bottom) with standard automatic preprocessing and algorithmic hemodynamic quantification.
Figure 3. Receiver-operator characteristic curves show performance for prediction of adverse aortic outcomes in BAV patients from CEMRA-based (right) or 4D flow MRI-based (left) hemodynamic parameter estimates; AUC values are shown as mean [confidence interval].