Research On Image Super-resolution And Its Applications In MRI

Spatial resolution of digital images is a common parameter to be considered in most imaging applications and an important index to evaluate image quality.Improving image resolution by upgrading imaging systems is limited by technical bottlenecks,application conditions and manufacturing costs.However,it is an effective and cost-efficient alternate to use signal processing and machine/deep learning methods to break through the inherent defects in hardware devices,which is termed as image super-resolution(SR).Image SR is a typical image restoration problem in the field of computer vision,which aims at overcoming the physical defects of imaging systems or inappropriate acquisition conditions,and recovering a high resolution(HR)image from one or more low resolution(LR)images.In recent years,it has become a major trend to build the nonlinear mapping between LR and HR images by introducing prior information from numerous external data through deep learning techniques.This dissertation carries out innovative study on image SR methods based on deep learning,and explores the potential applications in magnetic resonance imaging(MRI).The main contents of this work are as follows:(1)Considering that many SR models based on convolutional neural network(CNN)introduce residual learning to increase model scale and stabilize model training,so that the representational capacity of deep models is not fully exploited,we present a novel fully channel-concatenated network(FC~2N).In the FC~2N,all interlayer skip connections(ISC)are implemented by weighted channel concatenation.Because all the weighting factors of ISC are learnable,the model can adaptively decide the strength and number of ISC.This helps to make full use of model representational capacity,and is more consistent with the way brain neurons behave,thus more physiologically sound.To our best knowledge,our FC~2N is the first CNN-based SR model that does not use residual learning but still reaches network depth of over 400 layers,and the first CNN-based model that achieves advanced SR performance with less than 10M parameters after the EDSR model.(2)For the problem of gradient vanishing and model under-fitting caused by training large-scale CNN models using magnetic resonance(MR)images with simple structures and textures,a novel channel splitting network(CSN)is proposed in this work.The CSN model applies channel splitting to divide hierarchical features into different branches and merges them by the strategy of merge-and-run(MAR).Channel splitting divides feature maps into different clusters to deal with these sub-features discriminatively,while MAR can promote information sharing and integration among different branches by periodically combining hierarchical features on different branches in a residual manner.Experiments show that our CSN model not only surpasses traditional MR image SR methods,but also has significant performance advantages over other advanced CNN-based SR models.(3)In view of the problem that the local parallel module in the CSN model limits the network depth and model representational capacity,we further propose a channel splitting and serial fusion network(CSSFN)for MR image SR tasks.The proposed CSSFN model also groups hierarchical features by channel splitting,but feature integration is achieved by a serial structure similar to dense connected network(DCN).In this way,one can keep the model's ability to discriminate features on different channels,and expand its representa-tional capacity by increasing network depth.Extensive experiments show the performance advantages of our CSSFN to other advanced methods including the CSN.(4)Considering that large-scale CNN models require more computing and storage resources,which are unfriendly to real MRI scenarios,we present a new lightweight lateral inhibition network(LIN).According to the lateral inhibition effect in neurobiology,LIN model assumes that there exists mutual inhibition between adjacent neurons,and conducts explicit inhibitory regulation on hierarchical features by simulating the Hartline–Ratliff Equation.To further boost the performance of lightweight models,we propose to integrate dilated convolutions with different dilation rates to extract shallow features,which favor to extract features under receptive fields with different sizes and provide more abundant and effective evidence for subsequent inference.Experimental results demonstrate that the proposed LIN model achieves excellent SR performance with a small number of model parameters,exhibiting good practicability.(5)By feat of image SR framework based on deep learning and transfer learning,we explore other applications of related regression models in MR image through appropriate network adjustment and model hyper-parameter setting,including Gibbs-ringing artifact suppression(GAS),automatic windowing(AW)and age predication(AP)etc.Numerous experiments illustrate that CNN-based regression models can be easily applied to process these tasks on MR images,and perform obviously better than traditional methods.