Direct application of deep learning for diffuse optical tomography

Diffuse optical tomography (DOT) is a non-invasive imaging technique that uses near-infrared light to visualise tissue properties by reconstructing optical characteristics in scattering media. Time-domain DOT (TD-DOT) enhances this process by measuring photon time-of-flight, providing depth-resolved data but at a high computational cost. This study investigates the potential of machine learning (ML) for accelerating TD-DOT image reconstruction using simulated data. A neural network with an autoencoder architecture and a 3D U-Net refinement stage was trained on simulated datasets. The model achieved 96.4% precision in detecting the inclusion positions and reconstructed 3D optical properties in 0.25 seconds, with root mean squared errors of 20.1% for voxelwise absorption and 17.2% for scattering coefficients. These results demonstrate ML’s feasibility for fast, accurate TD-DOT reconstruction, highlighting future prospects for real-world validation and improved data compression techniques.