Cerebrospinal fluid (CSF) flow is crucial for clearing metabolic waste from the brain, a process whose dysregulation is linked to neurodegenerative diseases like Alzheimer's. Traditional approaches like particle tracking velocimetry (PTV) are limited by their reliance on single-plane two-dimensional measurements, which fail to capture the complex dynamics of CSF flow fully. To overcome these limitations, we employ artificial intelligence velocimetry (AIV) to reconstruct three-dimensional velocities, infer pressure and wall shear stress and quantify flow rates. Given the experimental nature of the data and inherent variability in biological systems, robust uncertainty quantification (UQ) is essential. Towards this end, we have modified the baseline AIV architecture to address aleatoric uncertainty caused by noisy experimental data, enhancing our measurement refinement capabilities. We also implement UQ for the model and epistemic uncertainties arising from the governing equations and network representation. Towards this end, we test multiple governing laws, representation models and initializations. Our approach not only advances the accuracy of CSF flow quantification but also can be adapted to other applications that use physics-informed machine learning to reconstruct fields from experimental data, providing a versatile tool for inverse problems.
Read full abstract