Research On Reconstruction Method Of Tumor Microvascular Based On Deep Convolutional Neural Network

Mesoscopic fluorescence molecular tomography(MFMT)can perform high-resolution(micron-level)non-contact imaging of thick tissue(millimeter-level depth)with a large field of view by looking for the relationship between the fluorescence signal measured on the tissue surface and the fluorescence parameters in the body(i.e.sensitivity matrix),which fills the microscopic and macroscopic vacuum areas of imaging spatial resolution well.However,the detected fluorescence signal has been scattered by biological tissues for many times,forming a nonlinear relationship with the sensitivity matrix,making the problem of reverse reconstruction an ill-posed problem.Moreover,the large-scale sensitivity matrix increases the computational complexity of the reconstruction algorithm.The specific optimization algorithm is needed to reconstruct the three-dimensional distribution of fluorescent markers in biological tissues.Therefore,to effectively promote the pre-clinical study of MFMT and make it better applied to early detection and treatment of tumors,this paper focuses on how to achieve "high efficiency" and "high precision" reconstruction,and proposes a hybrid strategy for microvascular reconstruction based on deep convolutional symmetric network(DCSN)and its improved network(doub-skip dense network)combined with simultaneous algebraic reconstruction technique(SART).Utilizing deep learning techniques that are widely concerned by current scholars to accelerate the solution of inverse problem while retaining the physical nature of the conventional reconstruction steps.The innovative points and main research work of the thesis include the following:(1)To obtain a strong mapping relationship between the sensitivity matrix and the fluorescence measurements of the tissue boundary,combining the two to form an augmented matrix as the training data.The low-dimensional spatial expression of the augmented matrix is learned to get the fusion feature of the sensitivity matrix and the fluorescence measurements.Starting from the data itself,to a certain extent,the purpose of accelerating the reconstruction of fluorescent markers within the tissue can be achieved.(2)The large-scale data obtained by multi-angle projection aggravates the burden of fluorescent molecular reconstruction,and reduces the reconstruction speed.In view of this,a hybrid reconstruction strategy of deep convolutional symmetric network combined with SART is proposed in the paper.With the help of the powerful feature extraction ability,nonlinear mapping characteristic and information integration ability of the convolutional neural network,completing the effective compression of large-scale training data,and realizing the rapid and accurate reconstruction of the distribution of fluorescent markers within the tissue by the SART algorithm.Designing an arc-shaped symmetric network that is suitable for encoding high-dimensional fluorescence data as the backbone architecture,and configuring the convolutional layer structure for specific training data.To reduce the parameters and increase the fitting ability of the network,multiple 3×3 small convolutional kernels filled with "valid" are used to complete the convolutional operation,and only one pooling operation is performed,so that the training data can be compressed as expected,and minimize the influence of redundant information on reconstruction results.Through the comparison results of the in silico,the processing performance of the DCSN model is demonstrated,and it is verified that the hybrid strategy has a better processing effect and faster reconstruction speed than the conventional method.(3)To strengthen the integration capability of the network for shallow and deep features,further promote the convergence speed,and improve the reconstruction effect of the hybrid reconstruction strategy.On the basis of DCSN,the paper designs and implements a microvascular reconstruction algorithm based on deep convolutional doub-skip dense network(doub-SD).The architecture of doub-SD introduces a local doub-skip residual structure and a global dense interactive connection,making full use of low-dimensional hierarchical features.Moreover,adding a 1×1 convolutional bottleneck layer increases the network depth and reduces the burden of weight parameters to obtain smaller and more useful data information with a more optimized learning method.By analyzing the effect of accelerating reconstruction of the in silico and the synthetic data of the vascular tree,the superiority of the improved network is verified.The reconstruction speed is increased by about 1.5 times,and the calculation time is reduced to 2/3 of the original.