Super-resolution Reconstruction Of Medical Images Based On Deep Learning

As we all know,nowadays,Medical imaging mostly relies on high-tech imaging equipment.The doctor mainly uses medical image to diagnosis patient’s disease and find out diseased biological tissue.Therefore,It is of great practical significance to reconstruct medical images with super-resolution and improve the resolution of medical images.Our master thesis uses deep learning to reconstruct CT medical image,assist doctors in the detection of diseased organ targets,decrease error rate.Medical imaging is limited by imaging principles,imaging equipment,and patient safety.During the imaging process,which will result in low resolution of medical im-ages.However,changes in hardware conditions are costly and have many limitations.In order to solve this problem,doctors can diagnose patients more quickly and accurately.We use software methods,use deep learning techniques.This master thesis summarizes and studies the classic deep learning models in the super-resolution field and the super-resolution models in the medical image field,and proposes two improved deep learning models-multi-scale joint convolutional neural network and multi-level multi-scale The residual convolutional neural network is used for super-resolution reconstruction of med-ical images.And we prduce CT medical data to test our SR model.In the mean time,we also research the basic knowledge in the medical image super-resolution field,summary the evaluation index.As we all know,The medical image is important and particular,the loss of image detail information may lead to doctors’misdiagnosis of patients.Therefore,in order to fully extract medical image features,we propose a multi-scale joint convolutional neural network.Unlike the previous SR model that uses a single convolution kernel for feature extraction,we use convolution kernels with different receptive fields for feature extrac-tion.At the same time,in order to reduce the amount of parameters,two stacked small convolution kernels are used instead of big convolution kernels to form a multi-scale joint block.In order to reduce the amount of parameters,a recursive network structure is used to build the network.At the same time,multi-channel global residual learning is introduced to prevent the model from wasting time on training of low-frequency information and identity information,accelerate network convergence and ensure the normal transmission of characteristic information.The multi-level multi-scale residual network is proposed to improve the network re-construction performance and stack more neural modules.We integrated the residual net-work into the multi-scale joint block to form a multi-scale residual block,In the mean time,At present,most SR models in the field of medical image usually use one-time magnifica-tion,and the quality of reconstructed image is not high under large magnification factor.So,we incorporated the idea of stepwise sampling of the Laplacian pyramid network into the model,and built a multi-level output network to form a multi-level multi-level output net-work.Scale residual network.Enhance the flexibility of the network.Other than that,the quality of the reconstructed image is enhanced under large magnification factorFinally,we obtain the original abdominal CT medical images from the TCIA website and make a super-resolution data set,and verify our deep learning SR model on the abdom-inal CT medical image data set for comparative analysis.Verifies the effectiveness and superiority of our proposed deep learning SR(Super resolution)model for super-resolution reconstruction in the field of medical images...