Machine learning techniques for predicting microvascular network hemodynamics

Existing techniques to image microvascular networks in vivo do not allow a direct and accurate measurement of several hemodynamic variables such as the wall shear stress (WSS). They cannot also provide a full 3D characterization of blood velocity and red blood cell (RBC) concentration profiles in every vessel in a microvascular network. Such physiological parameters however play critical roles in blood flow regulation, disease progression, angiogenesis, and hemostasis. Theoretical models of network blood flow, often used for hemodynamic predictions in experimentally acquired images of the microvascular network, cannot provide the full 3D blood velocity, RBC concentration and WSS profiles as these models assume one-dimensional flow governed by Poiseuille’s law. In contrast, high-fidelity computational models that simulate flow of blood as a suspension of deformable RBCs and model the 3D vessel geometry can readily predict such physiological variables with high accuracy. Such high-fidelity computational models have recently been applied to predict hemodynamic variables in microvascular networks in silico composed of many vessels and vascular junctions. Such models, however, tend to become computationally expensive with increasing size of the network; therefore, they are not feasible for use in large networks at organ scale. To overcome this limitation, we developed a machine learning (ML) approach that can predict hemodynamic variables in 3D and with high accuracy. The ML models are based on artificial neural network and convolution-based U-net models. Using experimentally acquired images of the microvascular networks, inputs at a few inlets, and a “bank” of database that we have developed, these ML models can predict flow and RBC distribution in all vessels in the network, 3D blood velocity and RBC concentration profiles over the cross-section of every vessel, 3D WSS profiles, and time-dependent variations in blood flow in any vessel. The prediction from the ML models compare very well against the true data. Also, the prediction time using ML is significantly less than that of the high-fidelity models. The current ML models, which to our knowledge are the first of their kind, are highly promising for image-based predictions of sub-cellular resolved capillary hemodynamics in organ-scale networks. This study paves the way for AI to make hemodynamic predictions in organ-scale capillary vessel networks while retaining the sub-cellular scale details and overcoming the limitations of the in vivo imaging techniques, with potential applications in hematological and microvascular disorders, angiogenesis, and vascular-mediated drug delivery. Funded by NIH (R01-EY033003) and NSF (CBET 2302212). This is the full abstract presented at the American Physiology Summit 2024 meeting and is only available in HTML format. There are no additional versions or additional content available for this abstract. Physiology was not involved in the peer review process.