Research On Fluorescence Tomography Method Based On Stacking Self-encoding

Fluorescence Molecular Tomography(FMT)as a non-invasive imaging technique has been the focus of biomedical research.FMT can specifically recognize tumors at the molecular level and can be applied to early tumor screening and treatment.However,the inverse problem of FMT is ill-conditioned and ill-posed.The traditional algebraic iterative method is slow in calculation and large in error,which makes it difficult for FMT to enter clinical applications.Therefore,the performance improvement of FMT imaging reconstruction algorithm has been one of the research directions of researchers in this field.In this thesis,the reconstruction process of FMT is optimized based on the Stacked Auto-Encoder(SAE)neural network model.The main research contents are as follows:(1)We constructed the SAE deep neural network model and obtained the fluorescence yield images of heterostructures at different locations in a two-dimensional circular domain simulation model.A total of 400 heterogeneous models with radius of 2 mm,3 mm,4 mm and 5 mm were constructed in the circular domain model.The mixed data set is composed of different Gaussian random noise levels and original pure data.According to the noise level(30 dB,35 dB,40 dB),it is divided into three groups,each group of 800 data samples.750 of them were randomly selected as the training set of modeling,and the remaining 50 as the prediction set.Comparing the quality of the reconstructed images based on SAE and traditional algebraic reconstruction technique(ART),the results show that the reconstruction method of SAE has better noise resistance and small heterogeneity recognition than the traditional ART reconstruction algorithm.Even when the radius of the heterogeneity is 2 mm and the signal noise ratio is 30 dB,the reconstruction method of SAE can still obtain clear and accurate images.(2)In order to further optimize the SAE network method and improve the quality of the reconstructed fluorescence yield images,the SAE network model structure is systematically studied in this thesis.The mean square error of reconstruction results is used to evaluate the performance of network reconstruction.The results show that when the number of double hidden neurons increases,the mean square error of network structure is smaller than that of decline mode.(3)In order to further optimize the model prediction effect based on SAE method,a visualization method of label distribution of data sets is proposed.The imaging reconstruction is performed for the different distributions of the five different training data sets and test data sets.The reconstruction results show that the relative relationship between the test data labels and the training data set label range has an important impact on the reconstruction quality of the model.This data label visualization method plays an important role in the selection of modeling data sets and error analysis of prediction data.The research in this thesis shows that FMT imaging based on SAE can improve the quality of the reconstructed images compared with traditional ART algorithm.The quality of the reconstructed images can be improved by optimizing the distribution of hidden layer neurons in the network structure and selecting data sets.