In this paper, we propose a novel 3D architecture called local Fourier neural operator (LoFNO) to improve hemodynamic analysis, which is essential for predicting cerebral aneurysm rupture and setting treatment directions. LoFNO overcomes the limitations of low spatio-temporal resolution and signal-to-noise ratio of MRI hemodynamic images, and predicts wall shear stress (WSS) directly from clinical image data. It improves structural recognition of irregular and unknown geometric structures by incorporating Laplacian eigenvectors as geometric prior information, and performs robust upsampling using enhanced deep super-resolution network (EDSR) layers. Combining geometric prior information and neural operator framework, LoFNO denoises and spatio-temporally upsamples blood flow data, achieving superior speed and WSS prediction performance than interpolation and other deep learning methods, thereby enabling more accurate cerebrovascular diagnosis.
