The Role of Spatial Preprocessing in Deep Learning-Based DOT

Diffuse Optical Tomography (DOT) is a non-invasive medical imaging technique that utilizes near-infrared light to study the optical properties of tissues. Recently, deep learning has gained popularity as a reconstruction method to solve DOT. However, despite its success, previous studies only reconstructed semi-homogeneous breasts with an absorption coefficient resolution of 2e-3 1/mm. In this paper, we propose a novel preprocessing method that considers the spatial correlations between different measurements to improve the reconstruction accuracy. Our algorithm is applied on a non-homogeneous breast phantom with absorption coefficient resolution of 5e-7 1/mm to reconstruct its optical properties. We compare our algorithm performance with and without the preprocessing step and to a SOTA analytical inversion technique. The proposed method is able to reduce the RMSE by more than 70% (0.44 to 0.11) and increase the contrast ratio by almost an order of magnitude (0.09 to 0.79).