Effect of deep neural network structure on the accuracy of NIR fluorescence molecular tomography reconstruction

To overcome the ill-conditioning of the NIR fluorescence molecular tomography (FMT) inverse problem, neural networks are commonly used for reconstruction to improve the accuracy and reliability of imaging. This paper aims to investigate the impact of different neural network structures on the reconstruction performance of FMT for improved effect. In this study, the finite element solution of the Laplace-transformed time-domain coupled diffusion equation serves as the forward model for FMT, an improved stacked autoencoder (SAE) network is used and applied to FMT reconstruction. In the study, the SAE was set as a four layers network model structure, of which two layers were used for the hidden layer of the network. When the number of neurons in hidden layer 1 is smaller than hidden layer 2, the network is referred to as a decreasing network structure, and vice versa for an increasing network structure. The input data to the network consists of surface fluorescence intensity values collected by detectors around the heterogeneity. The output data of the network consists of fluorescence intensity values on partitioned nodes obtained through finite element method (FEM) partitioning. The experimental results demonstrate that the increasing network structure exhibits better imaging accuracy, fewer artifacts, and a more stable network model in FMT reconstruction. Through this study of the impact of SAE network architecture on FMT reconstruction, we have identified the optimal network model, which holds significant guidance for the application of neural networks in the field of FMT.