Research On Deep Learnins Reconstruction Method For Tumor Mesoscopic Fluorescent Targets

Mesoscopic fluorescence molecular tomography(MFMT)is an optical molecular imaging technique at mesoscopic scale.It fills the vacuum zone of fluorescence molecular imaging and optical microscopic imaging at mesoscopic scale(10μm-1000μm)and can obtain the 3D distribution of fluorescence targets in biological tissue with a resolution about100μm.However,the reconstruction of fluorescent targets with a depth of several millimeters is essentially an ill-posed and ill-conditioned inverse problem and is susceptible to noise interference.Additionally,the dense spatial sampling required for high-resolution3 D reconstruction and the fine discretization of imaging targets make the sensitivity matrix columns highly correlated,while also increasing the consistency density of the measurements,which not only increases the memory footprint,but also increases the reconstruction time of traditional classical reconstruction algorithms and even prevents advanced algorithms from obtaining the best solution.Therefore,this paper proposes two reconstruction algorithms for the above problems: accelerated reconstruction algorithm of fluorescent target based on compressed sensing and convolution sparse autoencoder(Sparse Net)and highprecision 3D reconstruction algorithm of fluorescent target based on denoising convolution autoencoder(De Cnn Net).The specific research works are summarized as follows:(1)To achieve high-resolution three-dimensional reconstruction in mesoscopic fluorescence molecular tomography,dense spatial sampling and fine discretization of imaging biological tissue lead to the high correlation between rows and columns of sensitivity matrix,and at the same time,the measurement data obtained by MFMT imaging system is too dense.With the idea of compressed sensing theory,a new method based on compressed sensing and convolution sparse autoencoder is designed.For the algorithm,the multi-layer convolution of convolutional neural network is employed to decorrelate the sensitivity matrix,a sparse self-encoder network architecture is built to solve the problem of dense measurement data,and the traditional reconstruction algorithm is employed to reconstruct the fluorescent target with the help of preprocessed measurements and sensitivity matrix by the Sparse Net.The results of in silico and synthetic vessel tree experiment show that the designed strategy can greatly improve the reconstruction speed of traditional reconstruction algorithm.(2)To further improve the reconstruction accuracy of tumor mesoscopic target(microvascular)and suppress the large noise in mesoscopic fluorescence molecular tomography,a reconstruction algorithm based on convolutional denoising self-encoder is designed.In this algorithm,linear mapping module is used to solve the inverse problem of fluorescent target,and an U-shaped denoising self-decoder module is designed to denoise and sparse the initial solution.At the same time,to avoid the loss of important information(such as edge information and structure information of tumor microvessels)in convolution of neural network,data consistency layer(DCL)and transpose convolution are integrated into the network Layer.The in silico and synthetic vessel tree experiment show that this algorithm achieves the goal of accurate and direct reconstruction of fluorescent target with deep learning neural network,and has high robustness to noise.Compared with other algorithms,the reconstruction accuracy has been greatly improved.