PGDPNN: prior-information generation and distribution prediction neural network for fluorescence molecular tomography...

Fluorescence molecular tomography (FMT) is a promising and high sensitivity imaging modality that reconstructs the three-dimensional distribution of interior fluorescent sources. However, FMT reconstruction suffers from limited spatial resolution due to the simplifications in the forward model and the severely ill-posed nature of the inverse problem. In this study, we perform a clustering analysis using the radiomic features of the surface signal distribution. The FMT belonging to the clustering centers, along with their corresponding light source distributions, are defined as surface templates and source templates, respectively. Based on this, we propose a deep neural network-prior-information generation and distribution prediction neural network (PGDPNN)-for reconstructing the FMT source. PGDPNN consists of two key components: a spatial transformer network (STN) for prior source transformation and a V-Net for target source reconstruction. The STN estimates affine transformation parameters between the surface templates and the actual surface distribution (referred to as the target surface). These transformation parameters are then applied to the source templates to generate prior information as the light source distribution. Finally, the generated prior and the target surface are concatenated and fed into the V-Net to predict the final source distribution in FMT. Results shown in the numerical simulation experiments and in vivo experiment demonstrate that the proposed PGDPNN achieves excellent reconstruction accuracy and morphological recovery ability, with an average Dice coefficient of 0.7927 and a location error (LE) of 0.1363mm. The PGDPNN provides a novel research avenue for the rapid and accurate detection of tumors, enhancing the practicality of FMT in biomedical applications and contributing to the future development of FMT imaging technology.