Anatomy-derived 3D Aortic Hemodynamics Using Fluid Physics-informed Deep Learning.

Abstract

Background Four-dimensional (4D) flow MRI provides assessment of thoracic aorta hemodynamic measures that are increasingly recognized as important biomarkers for risk assessment. However, long acquisition times and cumbersome data analysis limit widespread availability. Purpose To evaluate the feasibility and accuracy of a generative artificial intelligence (AI) approach (fluid physics-informed cycle generative adversarial network [FPI-CycleGAN]) in quantifying aorta hemodynamics directly from anatomic input as an alternative to 4D flow MRI. Materials and Methods Patients were retrospectively identified from a dataset of clinical cardiothoracic MRI examinations performed between November 2011 and July 2020. All patients underwent aortic 4D flow MRI, which served as a reference standard for training and testing of FPI-CycleGANs. A three-dimensional (3D) segmentation of the aortic geometry was used as the only input to predict systolic aortic hemodynamics, with separate networks for bicuspid aortic valve (BAV) (994 in the training set and 248 in the test set) and tricuspid aortic valve (TAV) (419 in the training set and 104 in the test set). Voxel-by-voxel and regional analyses were used to quantify and compare (AI vs the reference standard, 4D flow) systolic velocity vector fields, peak velocity, wall shear stress (WSS), and classification of aortic valve stenosis. Results In total, 1765 patients (median age, 53 years [IQR, 41-63 years]; 1242 patients had BAV and 523 had TAV) were included. Mean AI computation time was 0.15 second ± 0.11 (SD), and total training was 1500 and 3600 minutes for the TAV and BAV networks, respectively. The FPI-CycleGAN predicted systolic 3D velocity vector fields accurately, with low bias (<0.01 m/sec) and excellent limits of agreements (±0.06-0.08 m/sec). For peak velocities and WSS, there was strong agreement between FPI-CycleGAN and 4D flow (r2 = 0.930-0.957 [P < .001], with relative differences of 6.2%-9.8%). AI accurately classified aortic valve stenosis severity in 85.8% of patients (302 of 352) (κ = 0.80 [95% CI: 0.71, 0.89]). The FPI-CycleGAN was robust to one- and two-voxel dilation and erosion (bias, -0.05 to 0.1 m/sec) and ±5° rotation (bias, -0.02 to 0.03 m/sec) of the input data. The application of the trained FPI-CycleGAN in an external test set with contrast-enhanced MR angiography (n = 60 patients) as AI input data demonstrated strong to excellent performance for peak velocities and WSS (r2 = 0.944-0.965 [P < .001], with relative differences of 6.2%-9.2%). Conclusion Aorta 3D hemodynamics can be derived from anatomic input in less than 1 second using an FPI-CycleGAN and demonstrate strong agreement with in vivo 4D flow MRI systolic hemodynamics. © RSNA, 2025 Supplemental material is available for this article.