FlowPredictor-Corrector: A Hemodynamics-Guided Two-Step Deep Learning Framework for Vascular Flow Prediction

Accurate hemodynamic assessment is of significant clinical value in the diagnosis, prognosis, and treatment planning of cardiovascular and cerebrovascular diseases. However, obtaining noninvasive, rapid, and high-fidelity hemodynamic information remains challenging. Invasive measurements carry risks, numerical simulations rely on complex modeling and high computational power, and while deep learning methods are efficient, they often lack physical constraints, leading to unstable performance in complex and diverse vascular geometries. To address this, we propose an innovative two-step framework called FlowPredictor-Corrector (FPC), which integrates point-cloud learning with physics-based hemodynamic correction to achieve a rapid mapping from vascular geometry to 4D hemodynamics. We constructed a dataset across three carotid artery anatomical regions (normal ICA C1 segment, normal carotid bifurcation, and stenotic carotid bifurcation) and conducted a systematic comparison with mainstream point cloud models (PointNet, PointNet++, Point Transformer). Our results demonstrate that the proposed method achieves high consistency with reference data for hemodynamic predictions ($R^{2} > 0.92$). Furthermore, it shows high agreement with computational fluid dynamics (CFD) and clinical ultrasound reports in stenotic grade classification based on internal carotid artery (ICA)/common carotid artery (CCA) peak systolic velocity (PSV) ($AUC = 0.947$). The model inference time was approximately 1.0 s per case. In summary, FPC not only balances efficiency with accuracy but also exhibits strong physical consistency and clinical interpretability, supporting further investigation toward rapid 4D hemodynamic prediction and potential clinical applications. To facilitate reproducibility and further research, the code and a representative test case are publicly available at https://anonymous.4open.science/r/FPC-B09E.