Acceleration And Super-resolution Design Of CT Image Reconstruction Based On Spark-GPU

Computer tomography(Computed Tomography,CT)is a technique that uses X-rays to penetrate a detected object to obtain projection data,and then performs a series of transformations on the projection data to obtain tomographic images.CT technology has matured after decades of development and is widely used in clinical diagnosis.The speed of image reconstruction is very important in the CT system.Researchers have been devoted to the accelerated research of CT image reconstruction.The development of computer hardware and big data technology has brought new opportunities and challenges to rapid reconstruction and image analysis.This paper combines the big data framework Spark and GPU to build a distributed computing cluster,in which to achieve accelerated reconstruction of batch CT images.The acceleration effect comes from two parts:1.The distributed cluster computing method can distribute the data to be processed to each computing node,so that the data that each node needs to process is reduced,so the total time is reduced.2.The addition of GPU can make use of its multi-threaded characteristics,which can greatly increase the speed and reduce the reconstruction time of a single image when dealing with CT image reconstruction such a computationally intensive and large amount of data.In this paper,real CT images are used to generate projection data,and then the projection data is used for reconstruction.The final experimental results show that the Spark-GPU platform designed in this paper has a significant acceleration effect on the filtering back projection algorithm and the simultaneous algebra iteration algorithm.In order to further optimize the efficiency and performance of the platform,this article analyzes the execution process of the program from four aspects:acceleration ratio,CPU usage,network bandwidth occupation,and memory usage,and optimizes the data transmission strategy.Medical images are limited to the characteristics of sensors and environmental conditions,and their resolution is not high,which has certain limitations on clinical diagnosis and subsequent image processing.Improving medical image resolution has always been the goal pursued by researchers.The development of sensor hardware is difficult to make breakthroughs in a short time,so improving the resolution of images through algorithms is an important way to solve this problem.Neural networks can automatically find optimal solutions and extract features.There have been remarkable achievements in the fields of classification,detection,and natural language processing Super-resolution models based on convolutional neural networks are constantly emerging with excellent results.This paper proposes a multi-scale wide activation super-resolution network,which combines the advantages of WDSR(Wide Deep Super Resolution)and MDSR(Multi-Scale Model Deep Super Resolution),and achieves double and quadruple super resolution at the same time by sharing the middle layer.In this model,small convolution kernels are used instead of large convolution kernels,and two methods such as smaller slices are used to achieve a good improvement in the effect.They are done with FSRCNN(Fast Super Resolution Convolutional Neural Networks)and traditional interpolation methods Compared.After the model training is completed,it is deployed on the Spark-GPU platform,using the master node to perform super-resolution processing,which not only can be connected with the CT image reconstruction process,but also achieves an increase in image processing speed,further proving the practicality of the platform.