A comprehensive approach to prediction of fractional flow reserve from deep-learning-augmented model.
Journal:
Computers in biology and medicine
PMID:
38194780
Abstract
The underuse of invasive fractional flow reserve (FFR) in clinical practice has motivated research towards non-invasive prediction of FFR. Although the non-invasive derivation of FFR (FFR) using computational fluid dynamics (CFD) principles has become a common practice, its clinical application has been limited due to the considerable time required for computation of resulting changes in haemodynamic conditions. An alternative to CFD technology is incorporating a neural network into the computational process to reduce the time necessary for running an effective model. In this study we propose a cascade of data-driven and physic-based neural networks (DP-NN) for predicting FFR (DL-FFR). The first network of cascade network DP-NN includes geometric features, and the second network includes physical features. We compare the differences between data-driven neural network (D-NN) and DP-NN for predicting FFR. The training and testing datasets were obtained by solving the three-dimensional incompressible Navier-Stokes equations. Coronary flow and geometric features were used as inputs to train D-NN. In DP-NN the training process involves first training a D-NN to output resting ΔP as one input feature to the DP-NN. Secondly, the physics-based microcirculatory resistance as another input feature to the DP-NN. Using clinically measured FFR as the "gold standard", we validated the computational accuracy of DL-FFR in 77 patients. Compared to D-NN, DP-NN improved the prediction of ΔP (R = 0.87 vs. R = 0.92). Statistical analysis demonstrated that the diagnostic accuracy of DL-FFR was not inferior to FFR (85.71 % vs. 88.3 %) and the computational time was reduced by a factor of approximately 3000 (4.26 s vs. 3.5 h). DP-NN represents a near real-time, interpretable, and highly accurate deep-learning network, which contributes to the development of high-performance computational methods for haemodynamics. We anticipate that DP-NN will enable near real-time prediction of DL-FFR in personalized narrow blood vessels and provide guidance for cardiovascular disease treatments.