Research On Medical Computed Tomography Image Super-Resolution Reconstruction And Application Technology

High-quality and high-resolution medical CT(Computed Tomography)images are of great significance in accurately reflecting the structure of biological tissues and improving the level of computer-aided diagnosis.In this paper,we mainly carry out related researches on super-resolution reconstruction methods of 2×2 acquisition mode and computer-aided diagnosis technology of bone mineral density(BMD)detection.On the one hand,the 2×2 acquisition mode has a faster scanning rate and relatively low radiation dose,but the resolution of acquired projection is low,and the reconstructed image has significant artifacts,which is not conducive to subsequent clinical diagnosis.On the other hand,the current screening capabilities of clinical BMD detection methods are limited,and these methods require higher hardware costs and a more cumbersome detection workflow.The vigorous development of deep learning related technologies has improved the effectiveness of CT image-based computer-aided diagnosis technology,which provide new opportunities to improve the comprehensive level of clinical BMD detection.Therefore,based on deep learning and the prior regularization related theories and technologies,researching efficient medical CT image super-resolution reconstruction methods and developing high-accuracy clinical BMD detection technology have important theoretical significance and practical application value.Aiming at the above problems,this paper focuses on three problems: CT super-resolution sinogram generation,super-resolution image reconstruction,and CT image BMD classification.The main research results are as follows:1.A CT super-resolution sinogram generation method based on generative adversarial network is proposed.In medical CT imaging,the resolution of projection obtained in the 2×2 acquisition mode is low,and the reconstructed image has significant artifacts,which will interfere with subsequent clinical image analysis.Aiming at this problem,this paper designed a generator based on U-Net and Res Net under the unique cycle consistent structure of Cycle-GAN to learn the mapping relationship between low-resolution and high-resolution sinograms,and designed a relative discriminator to enable the generator can more efficiently generate high-resolution sinograms.At the same time,this paper realizes the information interaction between the sinogram domain and the image domain by adding the filter back-projection module,and carries out the back-propagation of the cross-data domain error,so as to further enhance the network model to learn the details of the sinogram.The experimental results show that the proposed method can effectively improve the resolution of the sinogram obtained in the 2×2 acquisition mode and significantly suppress the artifacts caused by the LR attribute of the acquired projection.Compared with the reconstructed image based on the low-resolution sinogram,the error of the reconstructed image based on the proposed method is reduced by at least 70%.2.A CT super-resolution reconstruction algorithm based on block-matching and TV joint regularization is proposed.In the research of reconstructing the projections obtained in the 2×2 acquisition mode to obtain high-quality medical CT images,most of the existing algorithms only consider the degradation process of the image domain to establish the reconstruction model and do not sufficiently describe the projection degradation process in the 2×2 acquisition mode,which leads to its limited performance.Aiming at this problem,this paper proposed a CT super-resolution reconstruction model based on block-matching and TV joint regularization.The reconstruction model introduces the super-resolution sinogram generation method to overcome the nonlinearity of the projection degradation process,and enhances the constraints on the reconstructed image through the system matrix corresponding to the high-resolution projection and the low-resolution projection.In addition,the block matching sparse transformation is further used in the reconstruction model to mine the similarity of image blocks to improve the ability of image detail recovery,and total variation(TV)regularization is used to characterize the sparse characteristics of the image gradient-domain to enhance the noise suppression ability.Furthermore,the alternate minimization technique is used to efficiently solve the reconstruction model.The experimental results show that the proposed algorithm can better balance image noise suppression and detail information preservation,and further improve the reconstruction image quality in the 2×2 acquisition mode.Compared with the filtered back-projection reconstruction algorithm,the error of the reconstruction image based on the proposed algorithm is reduced by at least 50%.3.A CT image bone mineral density classification technology based on convolutional neural network is proposed.For the detection of BMD,the dual-energy X-ray absorption method and quantitative CT detection method are mainly used in clinical practice.These methods require additional auxiliary hardware and the cumbersome detection workflow,and it is difficult to screen for early bone mass loss without clinical symptoms.Aiming at this problem,inspired by the process of clinical radiologists to check CT imaging results to provide patients with diagnosis results,this paper firstly designed the MS-Net(Mark Segmentation Network)based on U-Net to realize the location and segmentation of regions of interest(ROI)in CT images.The MS-Net can narrow the range of attention to avoid the extraction of meaningless features.Then,this paper designed the BMDC-Net(BMD Classification Network)based on Dense Net for ROI feature extraction and analysis.The BMDC-Net can perform BMD classification to obtain the final qualitative diagnosis results.In addition,this paper used data augmentation technology to expand the dataset,and fully considered the rationality of the parameter scale in the network design to avoid network overfitting.The experimental results of clinical data show that the proposed method can report BMD qualitative detection results with high accuracy,and can be used as an auxiliary method to make routine medical CT examinations have the BMD detection function.Compared with traditional classification algorithms,this method can improve the accuracy of BMD classification by nearly 25%.