Monotone Accelerated Proximal Gradient Network For Bioluminescence Tomography Reconstruction

Bioluminescence tomography (BLT) is a noninvasive technique designed to enable three-dimensional (3D) visualization and quantification of viable tumor cells in living organisms. However, despite the excellent sensitivity and specificity of bioluminescence imaging (BLI), BLT is limited by the photon scattering effect and ill-posed inverse problem. To overcome this problem, regularization algorithms have been widely studied and achieved impressive results. Recently, reconstruction approaches based on deep learning have shown great potential in optical tomography modalities. However, the parameter selection of the regularization algorithm and the poor interpretability of deep learning methods have become the key factors to affect the reconstruction results and hinder its applicability. In addition, the spatial relationship between adjacent elements in the BLT data is a classical non-Euclidean data. Therefore, to mitigate the effects of this problem, in this paper, we proposed a novel Monotone accelerated proximal gradient network (MAPG-net) for bioluminescence tomography reconstruction by combining the advantages of the regularization method and Graph attention (GAT). The MAPG-net naturally inherits the solution constraints from the regularization-based methods framework, thus enhancing the stability and interpretability of the network. The numerical experiments confirmed that the proposed network has excellent performance.