Fast and Accurate Photon-Transport Modeling based on Foundation-Model-Encoded Implicit Neural Surrogate towards Optimized Near-Infrared Brain Stimulation

Abstract

Clinical treatment planning for near-infrared (NIR) neuromodulation requires patient-specific dosimetry to optimize light fluence (LF) delivery to cortical targets. The gold-standard Monte Carlo photon-transport forward solver is accurate but computationally expensive and non-differentiable for personalized inverse design across subjects. Here, we present PhomiNeuro, a foundation-model-encoded differentiable surrogate for time-resolved LF modeling and pathology-guided inverse design. A pretrained 3D medical imaging foundation model (VISTA3D) was domain-adapted to 285 training head models annotated with optical properties, and then coupled to an implicit neural representation (INR) that predicts LF at arbitrary spatiotemporal coordinates. Regularized by a time-dependent diffusion-equation residual, it demonstrated superior fidelity in 84 held-out participants, and ablations identified VISTA3D-derived anatomical priors as dominant and physics regularization as complementary. PhomiNeuro enables fast and robust queries of LF and its gradients with respect to illumination parameters for targets guided by individual structural magnetic resonance imaging and amyloid positron emission tomography, achieving 466-fold per-iteration speedup and 167-fold end-to-end optimization speedup. Explanatory analyses of LF delivery variability, including sex differences and amyloid status, identified structural variations as the primary drivers. These results position PhomiNeuro as a highly extensible translational framework toward personalized treatment planning and digital twin development in precision neuromodulation.