Development of a neural network predicting signals for time-domain diffuse optical tomography

Abstract Time-domain diffuse optical tomography (TD-DOT) is a powerful method for diagnosing anomalies in biological tissue such as brain hemorrhage and tumor. However, numerical simulations for TD-DOT require exploring a large number of parameter combinations and demand substantial computational resources. To address this challenge, we develop a neural network (NN) that can rapidly infer time-resolved signals from given tissue parameters. A high-quality training dataset for the NN is generated using ray-tracing-based radiative transfer simulations for 640 different absorber parameter combinations. Using the simulation data, we utilize NN to construct an emulator reproducing time-resolved signals for any parameters not used in the training data. We train two NN models with different training datasets: one with Gaussian noise added and the other without Gaussian noise. The NN trained with noisy data demonstrates superior performance, accurately reproducing time-resolved signals for unseen parameters. Its errors remain comparable to the noise level in the training data, highlighting strong robustness and generalization capability. Each inference takes only $$\sim 2\times 10^{-3}$$ seconds, which is $$>10^6$$ times faster than a direct radiative transfer simulation. This drastic speedup suggests the potential for efficient inverse problem analysis and application in real-time clinical diagnosis.