PhomiNeuro

Transcranial photobiomodulation shines near-infrared light through the scalp toward a cortical target, but scalp, skull, cerebrospinal fluid and parenchyma scatter and absorb it so strongly that only a small fraction of the energy arrives — and how small depends on the individual's skull thickness, CSF anatomy and brain parenchymal fraction. The reference answer comes from Monte Carlo photon transport, which is accurate, slow, and non-differentiable. That last property is the binding one: a Monte Carlo solver can say how much light reaches a target if the source sits at Cz, but it cannot say which way the source should point except by enumerating placements and re-simulating each.

PhomiNeuro turns the forward solver into something you can differentiate through. A pretrained 3D medical imaging foundation model, VISTA3D, is domain-adapted to four-channel optical-property volumes and then frozen; its multiscale feature pyramid conditions a source-conditioned implicit neural representation that predicts log fluence at any continuous spatiotemporal coordinate. Because the whole chain is a neural function of the illumination parameters, the gradient of delivered dose with respect to beam elevation and azimuth falls out of automatic differentiation, and beam placement becomes gradient ascent rather than grid search. The frozen encoder also means one trained model is applied to new participants, instead of being re-fit per scene the way coordinate networks usually are.

The model was built by researchers at Peking University Health Science Center, Inselspital (Bern University Hospital), Huashan Hospital at Fudan University, the Chinese University of Hong Kong and ShanghaiTech University. An earlier preprint version of the same work, posted in July 2026 under the title "Fast and Accurate Photon-Transport Modeling based on Foundation-Model-Encoded Implicit Neural Surrogate towards Optimized Near-Infrared Brain Stimulation", carries neither the PhomiNeuro name nor the evaluation numbers reported here.

#Key Features

• Continuous, off-grid queries: fluence is predicted at arbitrary (x, y, z, t) coordinates rather than on the simulation grid, recovering time points absent from training and extrapolating beyond the trained beam elevations.
• Differentiable inverse design: gradients of delivered dose with respect to source elevation and azimuth come from automatic differentiation, giving a 466-fold per-iteration and 167-fold end-to-end speedup over Monte Carlo optimization.
• Pathology-guided targeting: amyloid PET SUVR maps registered to each participant's head model define a burden-weighted cortical target, so the beam is aimed at pathology rather than at a standard 10–20 electrode site.
• Physics regularization: a normalized residual of the time-dependent diffusion equation is penalized at deep-tissue collocation points, complementing the encoder's anatomical priors.

#Technical Details

The encoder is initialized from the VISTA3D NV-Segment-CTMR variant and adapted with LoRA (rank 32, α = 64) under a masked-autoencoder objective, using an eight-head spatial redundancy-dropout attention borrowed from the fMRI foundation model NeuroSTORM to enforce long-range dependence. It emits a five-scale pyramid with channel dimensions [48, 96, 192, 384, 768]. Features are trilinearly sampled at the query point, concatenated with Fourier positional encodings and source-relative geometry, and passed to a four-layer GELU MLP of widths 512, 512 and 256. Training used 434 head models from SHARM/IXI, BrainWeb, scatterBrains and OASIS-3, split at the participant level into 285 training, 45 validation, 20 development-test and 84 held-out participants, with reference fields from MCX at 810 nm, 5×10⁷ photons, and a 2 ns window in ten 0.2 ns gates. On the 84 held-out OASIS-3 participants (42 Alzheimer's disease, 42 matched cognitively unimpaired), it reached a gate-averaged R² of 0.7385 ± 0.0019 and RMSE of 0.6211 ± 0.0027 decades, ahead of seven grid-based and coordinate baselines. Ablations put the adapted encoder first: removing it drops development R² from 0.7632 to 0.6508.

#Applications

The immediate use is patient-specific dosimetry and treatment planning for near-infrared neuromodulation in Alzheimer's disease, where fixed protocols ignore the anatomical variation that governs how much light reaches cortex. In 14 amyloid-positive AD participants, optimized orientations raised target fluence a median 3.8-fold over the best standard 10–20 position while recovering 91% of the Monte Carlo optimum, with every proposed pose independently re-verified by simulation. The same pattern fits other optical dosimetry problems — photodynamic therapy, optogenetic stimulation and photoacoustic fluence correction all need fast gradients through a light-transport model.

#Impact

PhomiNeuro is an unusual use of a medical imaging foundation model: a backbone pretrained for segmentation is repurposed as an anatomical-optical prior for a physics surrogate, and the ablations identify that prior as the dominant contributor to cross-cohort generalization. Its limits are stated plainly by the authors — a single wavelength (810 nm), prescribed beam profiles, bounded orientation ranges, head models that omit hair and assign homogeneous literature optical properties per tissue, and bright-fluence regions where high-photon Monte Carlo stays closer to the converged reference. The advantage claimed is continuity, speed and differentiability, not universal accuracy over Monte Carlo. The work is a preprint awaiting peer review, and the released encoder adapter inherits NVIDIA's non-commercial terms from the backbone it modifies.