1378 Determination of Blood Perfusion and Hemodynamics Using Physics-Informed Neural Networks for Use in Stroke Severity Assessment

INTRODUCTION: The current gold standard in prognostic stroke imaging is CT perfusion (CTP) imaging. This semi-quantitative modality approximates the infarct and penumbra regions within the brain from parametric maps, aiding in treatment decision making. While this gives a good indication of at-risk tissue during stroke assessment, there are many microvascular phenomena which are not successfully captured and reported during CTP imaging, but which contribute significantly to dead and at-risk tissue volumes in the brain. METHODS: A 1x1x1 mm microvascular model was used to simulate angiographic images within a computational fluid dynamics solver. This data was spatially binned into a 13x13x13 voxel volume, with a temporal resolution of 152 fps. This angiographic data, as well as the convection equation and a conservation of mass term, were used as constraints on the PINN. The ultrahigh resolution contrast distribution prediction from the PINN was spatially binned and compared to the original image data to assess adherence to observed phenomena, while the resultant velocity maps were reported as vector fields and utilized to calculate volumetric flow rates throughout the microvascular volume. RESULTS: The contrast distribution obtained from the PINN compared well to input angiographic data, demonstrating the PINN’s effectiveness at improving spatial fidelity. Additionally, the velocity fields captured the arterial inflow and venous dispersion of contrast, which aided in accurate calculation of volumetric flow rates throughout the volume. CONCLUSIONS: The results indicate the effectiveness of PINNs in achieving super-resolution from sparse data, as well as a quantitative measure of hemodynamics within a microvascular network. Compared to the relatively low-resolution, semi-quantitative estimations by CTP, this microvascular assessment of blood flow using PINNs is indicative of improved localization and accuracy of metrics which directly relate to ischemia during stoke.