| Single photon emission computed tomography (SPECT) is one of the nuclear medicalimaging technologies. It takes the advantages of functional imaging and dynamic imaging,in contrast with the other diagnosing technologies. Therefore SPECT plays an importantrole in the clinic diagnosis, especially in the heart diagnosis, brain diagnosis and the earlydetection of tumors.In SPECT, the reconstruction is based on the measurement of gamma photons emittedby the radiotracer. The number of gamma photons detected is proportional to the dose ofradiotracer, but the dose of radiotracer is limited in consideration of patients’ safety. Inorder to detect enough gamma photons, it requires staying for a certain time under everyangle, so a long time is taken to acquire SPECT projection data. During the scan, thepatients need keep motionless to reduce artifacts caused by patients’ movement, whichleads to discomfort. But if the projection data collected from few angle is enough toreconstruct the image, it will reduce the detection time substantially. Thus, it is veryimportant to research SPECT reconstruction from few-view projections.Artificial neural network (ANN) has the potential to achieve a superior performance inhandling nonlinear optimization problem. And the problem of image reconstruction can beconsidered as optimization problem, so it can also be solved by ANN systems. Thedynamics of the networks are the minimization of the cost function (or the energy function),which can be minimized by the supervised or unsupervised learning strategy.In this paper, we design a network for SPECT reconstruction from full-view andfew-view projections, and propose an optimized parallel computing method using CUDA.The network that we used to reconstruction is based on EM algorithm, which is combinedwith the total variation (TV) minimization to achieve the reconstruction from few-viewprojections. The results of the full-view and few-view reconstruction prove theeffectiveness of the network algorithm and the optimized parallel computing method. |
