TLS-Sim: An Open-Source Toolkit for Individualized Transcranial Light Stimulation Modeling with Deep-Learning-Accelerated Simulation
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Accurate modeling of photon transport through the heterogeneous tissues of the human head is important for individualized transcranial light stimulation (tLS). However, high-precision Monte Carlo (MC) simulations can be computationally demanding, and existing general-purpose simulation tools do not provide an integrated, brain-oriented workflow.
Methods
We developed TLS-Sim, a framework that integrates subject-specific head-model construction, multiple light-source configuration, GPU-accelerated MC photon simulation, visualization, and intracranial dosimetric analysis. Its central feature is a deep-learning-accelerated MC pathway, termed Flux-to-Flux, that estimates a high-photon-count energy-deposition field from a paired low-photon-count simulation. Sparse and full Monte Carlo fields were encoded using a three-dimensional variational autoencoder (VAE), and a three-dimensional velocity network was trained to transform the sparse latent representation toward the corresponding full-simulation representation.
Results
We demonstrate the end-to-end toolkit and the Flux-to-Flux (F2F) pathway on a representative individualized head model evaluated across three source configurations. Relative to the sparse simulation, the learned prediction improved agreement with the full Monte Carlo reference by approximately +4 dB in signal-to-noise ratio on average, with correspondingly higher structural similarity and lower mean absolute error.
Conclusions
TLS-Sim provides a modular, openly released computational framework for individualized tLS simulation and analysis. The Flux-to-Flux pathway is presented as an implemented capability that reduces the computational burden of high-photon-count simulation while improving spatial agreement with the full MC reference, most strongly at the low-fluence field boundaries where the sparse simulation is noisiest.
