A247 Use of a physics informed neural network (pinns) for in-vivo quantification of csf flow

<h3>Introduction</h3> Computational fluid dynamics (CFD) has long been used to model cerebrospinal fluid (CSF) flow in virtual environments, allowing exploration of complex interactions that are difficult to study in vivo. However, CFD faces significant challenges, including its reliance on non-intersecting meshes and the requirement to define boundary conditions—both of which are difficult to obtain in biological systems. Physics-Informed Neural Networks (PINNs), on the other hand, have emerged as a powerful alternative for modeling three-dimensional (3D) flow fields from sparse or incomplete data. <h3>Aim of Study</h3> To reconstruct 3D CSF flow fields using a PINN model trained on planar projection data extracted from CFD simulations in anatomically realistic geometries. <h3>Method</h3> A three-dimensional (3D) reconstruction of a canine SAS segment was used, produced from in-vivo real time imaging of the SAS using intravascular high-frequency optical coherence tomography (HF-OCT). CFD simulations of CSF flow within this segment were performed, and planar projection velocity data were extracted. A dedicated PINN model, developed in-house, was trained on this projection data to reconstruct the full 3D velocity field within the same <h3>Results</h3> The PINN model, initially trained on simple anatomical geometries, yielded results comparable to full CFD simulations. When applied to the complex SAS geometry, the PINN reconstructions achieved less than 10% error relative to CFD predictions. <h3>Conclusion</h3> These findings demonstrate that PINNs can accurately reconstruct complex CSF flow patterns from limited data, offering a promising alternative to traditional CFD by reducing the need for detailed boundary conditions and mesh construction in biological systems. <h3>Conflict of Interest</h3> No