Image by lucadp/Shutterstock

Researchers at University of Tsukuba have developed an artificial intelligence (AI) model capable of predicting light propagation in biological tissue in diffuse optical tomography (DOT), a non-invasive imaging technique for detecting abnormalities such as hemorrhages and tumors. The model performs these calculations in approximately 2 milliseconds, exceeding the speed of conventional simulation methods by more than one million times, paving the way for real-time diagnostic applications.

Tsukuba, Japan—Diffuse optical tomography (DOT), a safe, non-invasive imaging technique that utilizes near-infrared light, has attracted increasing attention as a diagnostic method for identifying abnormalities in conditions such as cerebral hemorrhage and malignant tumors. This technique detects internal abnormalities by illuminating biological tissue with near-infrared light, without causing radiation exposure or damage. However, high diagnostic accuracy depends on solving the radiative transfer equation that models light propagation within tissue. Since these numerical simulations can take several hours per calculation, using this method for real-time diagnosis remains challenging.

To eliminate computationally intensive simulations, this study introduces a neural network-based machine learning model that serves as an ultra-fast emulator. Trained on extensive simulation data, the model predicts time-resolved light signals detected at measurement points based on the location and size of an abnormal region. The model demonstrates robust generalization, accurately reproducing signals, even for unseen parameter combinations, with accuracy limited only by the noise level in the training data. Each inference takes approximately 2 milliseconds, representing a speedup of more than one million times compared with conventional simulation methods. This dramatic acceleration enables efficient exploration of the vast parameter spaces required for diagnostic analysis.

Furthermore, by combining the AI model with statistical sampling techniques, the model enabled the researchers to accurately estimate the location and size of abnormal regions from optical signals. These findings highlight this model as a promising foundational tool for real-time diagnosis of cerebral hemorrhage and tumors.

###
This research was supported by MEXT/JSPS KAKENHI Grant Numbers 21H04489 and JST FOREST Program Grant Number JP-MJFR202Z and Strategic Professional Development Program for Young Researchers, TRiSTAR fellow.

Original Paper

Title of original paper:

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

Journal:

Biomedical Engineering Letters

DOI:

10.1007/s13534-026-00578-9

Correspondence

Researcher HORIE Shu
Professor YAJIMA Hidenobu
Center for Computational Sciences, University of Tsukuba

Related Link

Center for Computational Sciences