Key points are not available for this paper at this time.
Hemodynamic analysis of the thoracic aorta is fundamental to the diagnosis and prognostic assessment of cardiovascular diseases. Although traditional computational fluid dynamics provides high-fidelity results, the requirements for complex mesh generation and significant computational overhead currently preclude its integration into real-time clinical workflows. To address these limitations, this study presents the dual-channel feature modulation network (DCFM-Net), a lightweight point cloud-based deep learning framework designed for the simultaneous prediction of aortic pressure and velocity fields. The architecture incorporates a distance field (DF) and positional encoding to refine geometric boundary characterization and enhance the representation of high-frequency spatial features, respectively. Furthermore, a feature-wise linear modulation mechanism is utilized during the feature fusion stage to adaptively modulate local features based on global geometric context, while a dual-branch decoder facilitates joint physical field prediction. Experimental results demonstrate that DCFM-Net reconstructs three-dimensional flow fields within milliseconds, achieving average errors of 4.50% ± 0.90% for velocity and 3.49% ± 1.66% for pressure. This approach outperforms several baseline neural networks and significantly improves computational efficiency. To enhance model transparency, the SHapley Additive exPlanations (SHAP) method is employed for interpretability analysis. SHAP quantifies the contributions of spatial coordinates and DF features to reveal the network's feature-utilization patterns. These importance trends are consistent with the spatial characteristics of the predicted flow fields. Overall, the proposed method offers a substantial improvement in computational efficiency without compromising accuracy, representing a promising data-driven surrogate modeling technique for the rapid assessment of hemodynamics within controlled computational environments.
Xiang et al. (Wed,) studied this question.