A deep convolutional neural network for diffuse correlation tomography

Near-infrared diffuse correlation tomography (DCT) is an emerging technology for tomographic imaging of blood flow index (BFI) in biological tissues through quantifying the light electric field temporal autocorrelation function. With the conventional approaches, proper reconstruction of BFI images is a challenging task from the limited DCT signals due to the severe imbalance between the optical measurements and the voxels to be reconstructed. In this study, we proposed a complete deep learning solution for DCT, including a dataset containing massive prior information for network training, a long short-term memory neural network for DCT signal denoising, as well as a deep convolutional neural network for mapping the DCT signals into the tomographic BFI images. The proposed deep learning solution was comprehensively validated through both computer simulations and phantom experiments, demonstrating its strong superiority over the conventional approach for precise and robustness reconstructions of the target BFI anomalies, with much better performance in reducing errors (i.e., the mean absolute error was reduced by 26.1 times) and preserving fine structure (i.e., the structure similarity index measure was increased by 12.8 times). The proper establishment of a deep learning strategy enables future exploration of the microvasculature blood flow mechanism on pathological tissues even from the limited DCT signals.