Research On GPU Acceleration And SDCA-ADMM Optimization Reconstruction Algorithm In Bioluminescence Tomography

Bioluminescence tomography(BLT)is based on the measurement of biological tissue fluorescence,to reconstruct the three-dimensional distribution of bioluminescent source.BLT including forward problem and inverse problem,The main problem that needs to be solved in the BLT forward problem is the accurate modeling and solution of the optical transmission mode.In several optical transmission models,the SPN model is the main object of the research.However,in the SPN model,the computational speed of the model is decreasing due to the rise of N.In the process of inverse problem reconstruction,by using the large-scale measurements problem can help improve the quality of reconstruction,but also consume plenty of time,so for large-scale measurements,we need do some research on fast reconstruction.In this thesis,we explore on the cost of SPN mode in the forward and reconstruction speed in the inverse of BLT.The main work includes:1)Solution and Implementation of BLT Forward Problem Based on GPU Parallel Computation.In the finite element method,the time analysis of the BLT forward process shows that the generating stiffness matrix and solving the linear equation are the most time consuming in the whole process,and these two parts are very suitable for parallel processing.Therefore,we use the SPN model as the optical transmission model and propose an accelerating strategy based on GPU/CPU dual platform in combination with GPU parallel processing capability.The forward process is decomposed into four parts as stiffness matrix generation,linear equation solution,data exchange and condition judgment.The first two parts are transmitted to the GPU for acceleration operation,and the latter two parts are handed over to the CPU for execution.The experimental results show that in this strategy,for different grid numbers,The SP7 model of the BLT forward overall acceleration ratio can reach a maximum of about 27 times.2)For the problem of BLT reconstruction speed of large-scale measurements,a BLT reconstruction algorithm based on the combination of alternating direction method of multipliers and stochastic dual coordinate ascent method is proposed.The objective function is decomposed into multiple sub-functions in the direction of a coordinate system by SDCA method,so that each coordinate system has a corresponding sub-function,and then the optimal solution of each sub-function can be calculated.In the iterative process,the sub-function in the corresponding coordinate system are selected according to the degree of the difference between the optimal solution and real value,and the usage rate of irrelevant measurement in the iterative process is reduced.So the convergence speed is improved and realize fast reconstruction.The experimental results show that the reconstruction speed is improved by about 5 times while ensuring the accuracy of BLT reconstruction image.