Fast Reconstruction Of X-ray Dynamic Micro-CT Based On GPU Parallel Computing

In recent years,with the wide use of the third generation synchrotron radiation facility,the photon flux density of X-ray imaging beamlines have been greatly improved,which provides the conditions for realizing X-ray dynamic micro-CT imaging.X-ray dynamic micro-CT imaging technology tends to have higher spatial resolution and faster time resolution.SSRF X-ray imaging group has developed a fast X-ray imaging detector based on long working distance microscope lens system and complementary metal oxide semiconductor(CMOS)digital high-speed camera.With this detector,the dynamic CT imaging time resolution can reach 25 Hz,and about 500 sets of CT data are generated in one day,The X-ray dynamic micro-CT user experiment of X-ray imaging beamline of SSRF will produce massive projection data in a short time.At present,CT reconstruction based on CPU reconstruction program is used in X-ray imaging beamline(BL13W),because CPU reconstruction program is serial computing,reconstruction time is long.Therefore,the speed of image reconstruction is much slower than that of original projection data,and there will be a large backlog of original experimental data.How to realize fast micro-CT reconstruction is a challenge for the wide application of dynamic micro-CT.Graphics processing unit(GPU)is widely used in the field of high-performance computing because of its multi-core architecture and large-scale parallelism.It can be used in CT reconstruction to realize the rapid reconstruction of X-ray dynamic micro-CT.In this paper,we use computer unified device architecture(CUDA)programming standard to parallelize the back projection operation of filtered back projection(FBP)CT reconstruction algorithm,and successfully implement the parallel calculation of CT reconstruction on GPU of NVDIA RTX2080.The results are as follows.1)Based on GPU parallel computing,the FBP reconstruction algorithm is accelerated.Compared with the CT reconstruction algorithm based on CPU serial computing,the GPU parallel reconstruction algorithm improves the reconstruction time by nearly 200 times,reduces the reconstruction time of a set of CT data(800 frames)from the original minute level to the second level,and significantly improves the image reconstruction efficiency of massive CT experimental data.2)It is found that in the CT reconstruction of GPU parallel operation and CPU serial operation,the image inversion algorithm can amplify the weak signal to a certain extent.In the process of CT reconstruction,for the tiny structure in the sample,the reconstructed signal is weak and the gray value of the reconstructed image is small.With the help of image inversion algorithm,the original pixel with small gray value can be amplified to a certain extent,so as to enhance the weak signal.The experimental results show that the contrast of reconstructed slices can be improved by using image inversion algorithm,and the brightness of three-dimensional reconstructed image changes more greatly,and the detail information is more obvious.3)The star artifacts in the process of back projection reconstruction are studied.Based on GPU parallel computing,different filters are selected to process the same set of projection data,and the performance differences of different filters are compared.The experimental results show that the three-dimensional CT reconstruction results with Hamming filter have less artifacts and better image reconstruction quality.4)The dynamic micro-CT experiment is designed and a 10 Hz time resolution dynamic micro-CT experiment platform is built to verify the CT reconstruction algorithm based on GPU parallel computing and Hamming filter.The experimental results show that the reconstruction time of 50 sets of tenebrio molitor bubble CT data is 5.93 min,while the reconstruction time based on CPU is 20.4 hours.By selecting the CT reconstruction data of bubbles at different time points,the dynamic change process of bubbles was observed in the time dimension,and the quantitative analysis of moving bubbles was made.After calculation,the bubble radius in tenebrio molitor was 0.624 mm,and the average moving speed of bubbles was 1.08 mm / s.5)At the same time of accelerating the CT reconstruction,the phase information of the sample is obtained by combining with the phase recovery algorithm,so as to accurately obtain the three-dimensional structure information of the sample.The experimental results show that the phase recovery algorithm can improve the image contrast of CT reconstruction results,and can clearly distinguish the organizational structure of tenebrio molitor.The pixel value of the local area of the image is very close.The internal organizational structure can be extracted and divided by ordinary image segmentation algorithm,which has important value for the analysis of the later image information.To sum up,based on GPU parallel computing,this paper develops a parallel program for the back projection process of FBP algorithm,and further optimizes the quality of image reconstruction by combining image inverse algorithm,filtering function and phase recovery algorithm.Finally,the research content is verified by designing dynamic micro-CT experiment and building fast X-ray imaging experimental platform.The experimental results show that the fast reconstruction of dynamic micro-CT is realized,the quality of image reconstruction is improved to a certain extent,and the problem of overstock of original experimental data faced by users of X-ray dynamic micro-CT experiment of SSRF is effectively alleviated.