Research On Optimization Of FMT Reconstruction Algorithm Based On Neural Network

Fluorescence molecular tomography(FMT),as one of the most promising imaging technologies,has gradually emerged in the field of molecular imaging.Because the physical nature of FMT imaging is based on diffused photon imaging,the similarity of its adjacent projections and the incomplete measured optical signals make FMT reconstruction highly ill-conditioned,limiting the clinical application of this technology.Therefore,the research on FMT reconstruction algorithms has been one of the research focuses of researchers.Traditional Algebra Reconstruction Technique(ART)has the disadvantages of slow reconstruction speed and large error.Deep learning algorithms do not need to specify the forward and reverse problems of FMT reconstruction,and fundamentally avoid errors due to inaccuracy of linear models.This article uses stacked auto-encoder(SAE)to implement FMT reconstruction,and optimizes the network to improve FMT reconstruction accuracy.The main research contents include:(1)The SAE model for FMT imaging is designed to reconstruct the fluorescence yield of heterogeneous bodies in different positions in a two-dimensional circular domain model.A two-dimensional circular domain model and the principle of FMT imaging are used to construct a SAE network model.Change the parameters such as the position and size of the heterogeneous body in the model,and complete the data set accumulation based on the radiative transfer theory and the finite element theory.The experiments use deep learning algorithms and traditional ART reconstruction at the same time,and compare the two.The results show that for the reconstruction of small target heterogeneous bodies such as 2mm and 3mm,the deep learning reconstructed image is clearer and higher resolution.(2)In order to optimize the deep learning reconstruction algorithm,the influence of the network structure on the reconstruction results was studied.The mean square error of the test data is selected as the evaluation index,and the results show that the reconstructed image of the incremental structural network model is clearer and more accurate.(3)Through the sensitivity analysis of the nodes,the key nodes are determined;the over-sampling technology is used to improve the signal-to-noise ratio of the key nodes and the accuracy of FMT reconstruction.The visual map was selected as the evaluation standard.The results show that oversampling of key nodes can improve the signal-to-noise ratio of the data and improve the accuracy of FMT reconstruction.The results show that the SAE-based FMT reconstruction algorithm can more clearly reconstruct the fluorescent heterogeneous body than the ART algorithm.In addition,this article combines a series of studies to optimize the network structure of SAE,enhance the signal-to-noise ratio of key node data,and achieve high-precision reconstruction of SAE-based FMT.