CorNet: A deep learning method based on physics-guided and attention mechanism for predicting flow field of coronary arterial tree

Replacing traditional computational fluid dynamics (CFD) with deep learning techniques has become a prevalent approach for studying blood flow and diagnosing diseases. Nevertheless, no neural network has been specifically designed for tree-like blood vessel structures that can effectively capture their bifurcation patterns and branching dependencies. Coronary neural network (CorNet) was proposed for predicting the pressure field and fractional flow reserve (FFRCorNet) of the coronary artery tree. The novelty of this framework lies in utilizing self-attention mechanisms to address long-term spatial dependencies within the vascular tree. Our dataset comprised 295 coronary arterial trees from 273 patients, each including point clouds reconstructed from medical images and fractional flow reserve values calculated by CFD (FFRCT). Physical constraints were incorporated to mitigate data sparsity and enhance the interpretability of the neural network. The pressure results predicted by CorNet are consistent with the pressure calculated by CFD (mean relative error = 3.96%). There is also a good consistency between FFRCT and FFRCorNet.Compared to invasive fractional flow reserve, which is considered the "gold standard", FFRCorNet demonstrates accuracy comparable to FFRCT (88% vs. 90%) while reducing computation time by several thousand-fold. Compared to previous studies, CorNet eliminates the need to identify specific lesion sites or manually extract geometric parameters of stenotic segments. This is the first computational method to predict hemodynamics in three-dimensional vascular tree structures using attention mechanisms within a deep learning model. We foresee that this framework will enable near-real-time flow field predictions for arterial trees and offer valuable insights for cardiovascular disease treatment.