Improved image quality of diffusion optical tomography through a deep learning based post-processing method

Diffusion Optical Tomography (DOT) is an emerging optical imaging modality with the advantages of being low-cost, non-invasive and non-damage-free, it has demonstrated great potential in differentiating benign and malignant breast tumors. However, due to the strong scattering of biological tissues and the limited number of boundary measurements, the DOT image reconstruction is ill-posed and ill-conditioned. Conventional DOT image reconstruction algorithms such as Tikhonov regularization, are easy to implement. Nevertheless, images with poor quality were usually achieved. In this paper, to improve the image quality of DOT, we develop a post-processing reconstruction algorithm based on deep learning. It has a type of U-net architecture, but with the low-quality image obtained with Tikhonov regularization and the acquired multi-wavelength optical signals as the input. Meanwhile, the output of the network is images of chromophores, including HbO, Hb and water. To train the proposed algorithm, we construct a dataset with 15,000 images for training, 6,000 for validation, and 3,900 for testing. Our results demonstrate that it has achieved good reconstruction results for the phantoms with single or double inclusions in terms of PSNR and SSIM. Compared to the Tikhonov regularization, the average PSNR of HbO, Hb and water are improved by 55%, 74% and 71%, respectively, and the average SSIM are improved by 11%, 76%, and 70%, respectively.