| Fluorescence molecular tomography(FMT)is one of the most successful optical molecular imaging techniques with many applications.It uses a specific molecule or cell labeled with a transmissive fluorescent molecular probe to observe the physiological and pathological changes in vivo at the cellular and molecular levels.Fluorescent molecular imaging technology has natural advantages in detection sensitivity,imaging depth,and spatial resolution,and the rapid progress in molecular probe research has made it have great potential for early detection of tumors.However,the inaccurate physical model description problem and the ill-posed problem in the inverse solution exist in the traditional FMT reconstruction technology,so it is not easy to obtain stable and excellent imaging quality.Therefore,FMT imaging technology has many worthy of further research and improvement.This thesis breaks through the tedious reconstruction ideas of traditional FMT,proposes two reconstruction network models,and introduces deep learning technology to make full use of a large number of different cases for learning to solve FMT reconstruction problemss.This method does not need to clearly define the forward and inverse problems of FMT reconstruction,nor does it require a priori knowledge of optical parameters;and,compared with the traditional FMT reconstruction algorithm,this method has great advantages in reconstruction accuracy and operation speed.The main research contents are as follows:1.An end-to-end 3D-Encoder-based decoder(3D-En-Decoder)reconstruction method is proposed.This method directly establishes a nonlinear mapping relationship from internal fluorescence distribution to boundary fluorescence signal distribution,that is,input to output.The encoder extracts the important information highly related to the class labels in the input data,removes the redundant content,and obtains the sparse representation of the projection image.Then,the decoder inputs the sparse feature representation of the projection data obtained by the encoder to the 3D decoder,using a series of deconvolution layers to predict 3D voxel output.The difference between the reconstructed image and the real fluorescence distribution map is learned through a content-based loss function,thereby establishing an accurate inverse problem representation function.2.A three-dimensional FMT reconstruction algorithm based on generating confrontation learning ideas and skip connection mechanism is proposed,the generator is 3D U-net and the discriminator is Res Net.This method uses a heuristic learning method when defining the loss function,so that the reconstructed image not only pays attention to the difference between the voxel level and the real image,but also the difference between a certain area in the image and the relationship between voxels and voxels.In addition,jump connections are added to both the generator and discriminator networks,and the input data is added to the network as a priori information,which solves the problem of gradient disappearance caused by the increase in the number of network layers,helps the back propagation of gradients,accelerates the training process of the network,and can pass more image detail information to enhance the influence of the feature map,so as to better reconstruct the image.This method utilizes the powerful representation capability of the generated adversarial network and the jump connection mechanism to further optimize and improve reconstruction details,improve reconstruction accuracy,and enhance reconstruction robustness. |
