Research On Super-resolution Algorithm Of Single 3D Magnetic Resonance Image

Magnetic resonance imaging(MRI)is a non-invasive,non-ionizing radiation,multi-parameter imaging technology.In clinical practice,high-resolution(HR)MRI images can provide clear anatomical details for the diagnosis and analysis of major diseases such as tumors.However,MRI has a long imaging time and is prone to motion artifacts due to involuntary patient movement during the scanning process.Although the scan time can be reduced by increasing the slice thickness of the MRI image,it will lead to partial volume effects and a reduction in the resolution of the MRI image.The blurred tissue boundaries in low-resolution(LR)MRI images limit its application in clinical disease diagnosis and analysis.Super-resolution(SR)is an image post-processing technology that can break through hardware limitations and improve the resolution of MRI images.In recent years,the powerful fitting ability of Convolutional Neural Network(CNN)has made a breakthrough in the research on SR of MRI images.However,the current CNN-based SR models have problems such as low feature utilization,high time complexity,poor visual perception quality of reconstructed images,and insufficient prior information mining of MRI images.In response to the above problems,this thesis conducts innovative research on CNN-based3 D MRI image SR technology,focusing on analyzing the model construction scheme of the MRI image SR method.We proposed a 3D MRI image SR method with high feature utilization,good visual perception quality,and image priors and explored its potential applications in tasks such as tumor segmentation.The main work and contributions of this thesis include the following aspects:In order to solve the problem that the existing CNN-based 3D MRI image SR models have high requirements for computation and memory,and are not suitable for MRI imaging scenarios with limited computing resources,a dense network-based fast SR algorithm(DNSR)for isotropic MRI is proposed.Different from the method that takes pre-interpolated MRI images as input,the DNSR algorithm takes original LR MRI images as input,which not only preserves the real data distribution of the original LR images but also effectively reduces the model computation.The DNSR algorithm continuously transmits features across layers through dense connections,which enhances the flow of information in the model and improves the utilization of features.Finally,the DNSR algorithm utilizes deconvolution layers to learn upsampling kernels to fuse multi-level features and generate HR MRI images,overcoming the dependence of pre-sampling methods on interpolation algorithms.Through a large number of ablation experiments,comparative analysis experiments,and application verification of tumor detection and segmentation,it is shown that the proposed DNSR method has good SR accuracy in single-scale,multi-scale and multi-modal MRI image SR tasks.At the same time,DNSR method has a small amount of calculation and memory requirements.MRI image SR methods based on image intensity constraints can only fit the mean value of multiple high-resolution images,which leads to MRI image edge blurring and being inconsistent with human visual perception.In order to solve the above problems,a generative adversarial network-based SR algorithm(GANSR)for isotropic MRI images is proposed.The proposed GANSR algorithm consists of a generator network that generates corresponding HR MRI images from LR MRI images and a discriminator network that distinguishes the generator-generated HR MRI images from ground truth HR MRI images.The GANSR algorithm not only uses the image intensity loss to constrain the content of the image,but also uses the idea of generative confrontation to further constrain the data distribution and texture details of the generated image,thereby improving the visual perception quality.Considering that different features and different feature regions have different contributions to the final generated results,in the design of the generator network,the channel attention and spatial attention mechanism are adopted to adaptively enhance the features and feature regions that are beneficial to improve SR performance.The experimental results demonstrate the effectiveness of the proposed channel and spatial attention mechanism.It is verified that the images generated by the GANSR algorithm are more in line with human visual perception on multiple public MRI datasets.Anisotropic 3D-MRI images have high-resolution layers,which can provide rich prior information for improving the resolution of MRI images.In order to solve the problem of insufficient prior information mining of MRI images by existing methods,an anisotropic MRI image SR algorithm based on multi-branch network(MBNSR)is proposed.The MBNSR algorithm transforms the problem of anisotropic MRI image SR into the problem of inserting additional slices into adjacent slices at in-plane.The MBNSR algorithm adopts a main branch to extract features from the LR slices,and the other two reference branches take adjacent HR slices as inputs,and extract high-frequency information beneficial to the target slices.Finally,the features of different branches are fused to provide rich edge and high-frequency information for SR reconstruction of target slices.At the same time,a joint loss based on image intensity and gradient is proposed to improve the texture details of the generated image.Experimental results show that the multi-branch network can effectively learn information beneficial to the target slice from adjacent HR slices.The MBNSR algorithm can more clearly restore the detailed contours of normal tissues and lesions in HR 3D MRI images.