CE-MRA-FLOWnet: Fast and Accurate Generative AI-Based Aortic Hemodynamic Mapping from Contrast-Enhanced MRA.

Abstract

Background: 4D flow MRI hemodynamic biomarkers (e.g. peak velocity and wall shear stress [WSS]) have shown promise for improved risk-stratification in patients with thoracic aortic disease (TAD). However, lengthy 4D flow scan times, complex data analysis and need for dedicated expertise limit clinical translation. In this study, we developed a fluid-physics informed deep generative adversarial neural network, CE-MRA-FLOWnet, to predict aortic hemodynamics directly from standard-of-care contrast-enhanced magnetic resonance angiography (CE-MRA) images.

Methods: We retrospectively identified 1392 patients (age: 52±12years, 1011 male, 954 with bicuspid aortic valve, BAV; 438 with tricuspid aortic valve, TAV) who underwent paired clinical 4D flow MRI and CE-MRA between 2011 and 2020 for suspected TAD. The CE-MRA-FLOWnet used CE-MRA data (1127 for network training, 265 for testing) to generate a prediction of aortic hemodynamics. 4D flow-measured aortic systolic 3D blood flow velocity vector fields served as ground truth. Analysis included comparison of AI-derived and ground truth aortic systolic peak velocity (PV) and WSS, the relative area of the AAo (%) exposed to elevated WSS, and aortic valve stenosis severity grading. The predictive value of CE-MRA-FLOWnet hemodynamics metrics was assessed in a subgroup of 133 BAV patients with known multi-year adverse outcomes.

Results: CE-MRA-FLOWnet training time was 8100mins; inference time per CE-MRA was 0.88±0.05seconds. AI-derived PV showed strong regional agreement and low biases with ground truth 4D flow (0.00-0.04m/s), and relative differences within 10.6%-12.7%. Aortic WSS also demonstrated minimal bias (-0.01-0.04Pa) and close alignment of the area of elevate WSS between AI and 4D flow (19.6±15.1% vs. 19.3±16.2%, p=0.84). AS severity was accurately classified in 88% of cases, with all grading errors limited to a one-grade difference (Kappa 0.84). ROC analysis showed that CE-MRA-FLOWnet derived hemodynamic metrics outperformed diameter alone for predicting adverse outcomes (PV AUC = 0.73-0.76; WSS = 0.83-0.88; diameter = 0.52-0.62).

Conclusions: CE-MRA-FLOWnet accurately predicts aortic hemodynamics in TAD patients using standard-of-care CE-MRA images. Our results demonstrate potential for clinical integration by providing physicians with near real-time hemodynamic data from widely available clinical CE-MRA images.

Clinical perspective: Approximately 3% of the population has or is at risk for thoracic aortic disease (TAD) which can result in significant complications including progressive aortic dilation and dissection. Historically, aortic diameter measured on CT angiography or MR angiography (MRA) has been the primary marker of risk and TAD patients undergo frequent surveillance imaging to evaluate aortic size and growth. However, it is well established that aortic diameter is an imperfect risk assessment tool and a large percentage of patients who are outside of guideline-endorsed diameter surgical thresholds can have aortic dissection. Recently, aortic hemodynamics measured with 4D flow MRI have shown promise for detecting patients who have higher aortic growth rates and 4D flow derived wall shear stress (WSS) has been linked to histopathologic damage to the aortic wall. Thus, aortic hemodynamic assessment could supplement aortic diameter to improve TAD risk-stratification. Unfortunately, 4D flow MRI is not widely available and can be difficult acquire, analyze, and interpret which has led to low utilization. In this study, we have developed CE-MRA-FLOWnet, a fluid-physics informed generative neural network that quantifies peak systolic aorta hemodynamics using only standard anatomic contrast-enhanced (CE) MRA as input. We found that CE-MRA-FLOWnet peak systolic velocity and WSS quantification is highly accurate compared to 4D flow MRI. We also found that these outputs can accurately grade aortic stenosis severity and are superior to aortic diameter for predicting adverse aortic outcomes. CE-MRA-FLOWnet could significantly expand access these important risk metrics to many more TAD patients.