Artificial intelligence-based surrogate modeling of hemodynamic indicators in abdominal aortic aneurysms using computational fluid dynamics

Abdominal aortic aneurysm (AAA) is a progressive dilatative disease of the abdominal aorta with high rupture-related mortality, emphasizing the need for early risk assessment. Conventional anatomical indicators, such as maximum diameter, are insufficient to characterize rupture risk, motivating the incorporation of hemodynamic parameters. Image-based CFD enables the estimation of these parameters but remains clinically limited by high computational cost and long processing times. To overcome these limitations, deep learning-based surrogate models such as MultiViewUNet have recently been proposed, though previous studies were constrained by relatively low spatial resolution and limited coverage of hemodynamic parameters. In this study, we propose MH-UNet, a multi-view, high-resolution surrogate modeling framework that extends the original MultiViewUNet to a higher input resolution and additional hemodynamic parameters. MH-UNet converts three-dimensional AAA geometries into two-dimensional images from multiple viewing angles and predicts TAWSS, OSI, RRT, and ECAP using data from 101 patients. Both single-inference and dual-inference strategies were implemented and compared. Single-inference consistently outperformed dual-inference across all parameters; the cascaded dual-inference scheme amplified errors through the nonlinear OSI and RRT formulations, resulting in negative R² values for OSI, RRT, and ECAP. Among the single-inference predictions, TAWSS was predicted most reliably (R² = 0.803), whereas OSI was the most challenging to predict. Agreement between single-inference predictions and CFD-derived results demonstrates the potential of MH-UNet as a rapid surrogate tool for estimating WSS-derived hemodynamic indices in AAA under fixed boundary conditions.