Interpretable model-driven projected gradient descent network for high-quality fDOT reconstruction

Abstract

In fluorescence diffuse optical tomography (fDOT), the quality of reconstruction is severely limited by mismodeling and ill-posedness of inverse problems. Although data-driven deep learning methods improve the quality of image reconstruction, the network architecture lacks interpretability and requires a lot of data for training. We propose an interpretable model-driven projected gradient descent network (MPGD-Net) to improve the quality of fDOT reconstruction using only a few training samples. MPGD-Net unfolds projected gradient descent into a novel deep network architecture that is naturally interpretable. Simulation and in vivo experiments show that MPGD-Net greatly improves the fDOT reconstruction quality with superior generalization ability.