Residual U-Net for accurate and efficient prediction of hemodynamics in two-dimensional asymmetric stenosis

Abstract

This study presents residual U-Net (U-ResNet), a deep learning surrogate model for predicting steady hemodynamic fields in two-dimensional asymmetric stenotic channels at Reynolds numbers ranging from 200 to 800. By integrating residual connections with multi-scale feature extraction, U-ResNet achieves exceptional accuracy while significantly reducing computational costs compared to computational fluid dynamics (CFD) approaches. Comprehensive evaluation against U-Net, Fourier Neural Operator (FNO), and U-Net enhanced Fourier Neural Operator (UFNO) demonstrates U-ResNet superior performance in capturing sharp hemodynamic gradients and complex flow features. For pressure prediction, U-ResNet achieves a normalized mean absolute error (NMAE) of 1.10%. Similarly, the performance of U-ResNet for wall shear stress (NMAE: 0.56%), velocity (NMAE: 1.06%), and vorticity (NMAE: 0.69%) consistently surpasses alternative architectures. Notably, U-ResNet demonstrates robust generalization to interpolated Reynolds numbers without retraining - a capability rarely achieved in existing models. From a computational perspective, U-ResNet delivers a 180-fold acceleration over CFD, reducing simulation time from approximately 30 minutes to 10 seconds per case. The model with non-dimensional formulation ensures scalability across vessel sizes and anatomical locations, enhancing its applicability to diverse clinical scenarios. These advances position U-ResNet as a promising auxiliary tool to complement CFD simulations for real-time clinical decision support, treatment planning, and medical device optimization. Future work will focus on extending the framework to three-dimensional geometries and integrating it with patient-specific data.