A Physics-Informed Neural Network Surrogate for Patient-Specific Hepatic Arterial Hemodynamics in Yttrium-90 Radioembolization: Network Architecture, Boundary-Condition Enforcement, and Data Efficiency

A novel approach has been developed to quickly predict the flow distribution of yttrium-90 microspheres during patient-specific trans-arterial radioembolization in liver cancer treatment. This physics-informed neural network (PINN) surrogate uses spatial coordinates to map velocity vectors and pressure fields. By leveraging a converged finite-element CFD solve with 105,284 nodes as reference data, the surrogate is trained on controlled data fractions, ranging from 0 to 0.20, to validate its performance.

The NVIDIA PhysicsNeMo framework was employed to compare different network architectures (MLP, Modified Fourier, Multi-Scale Fourier, and SIREN) under both soft and hard boundary condition enforcement. The results showed that the Multi-Scale Fourier network achieved the highest accuracy (R2(u) = 0.978) at a lower computational cost (300,000 steps) compared to other architectures, such as SIREN (R2(u) = 0.976) and MLP (R2(u) = 0.968).

The best performance was observed near a supervised data fraction of 0.01, with R2(u) values ranging from approximately 0 to 0.97-0.99. The best Multi-Scale Fourier model achieved R2(u) = 0.985 at the df = 0.20 checkpoint, matching CFD predictions with near-perfect accuracy. Once trained, the network can evaluate the entire 105,284-node field in approximately 2 seconds on a four-core CPU, making it a valuable tool for rapid predictions in patient-specific hepatic arterial hemodynamics during Yttrium-90 radioembolization.