| Magnetic Resonance Imaging(MRI)is an important medical imaging technique with non-invasiveness,non-ionizing radiation,and high contrast.However,its imaging speed is relatively slow,mainly limited by physical factors,including magnetic field strength and spin rate.Slow imaging speed has become a major factor affecting the development of MRI technology as it leads to long data acquisition times,which can cause discomfort to patients.In addition,during long acquisition processes,patients may involuntarily move,and these movements can introduce severe motion artifacts in the magnetic resonance images,thus affecting the diagnosis of MRI.Therefore,improving the imaging speed of MRI has great clinical value and economic significance.Inspired by super-resolution(SR)methods in the natural image field,researchers have applied SR methods to the MRI field to accelerate MRI imaging.Specifically,low-resolution(LR)images are obtained through fast scanning,and then high-resolution(HR)images are obtained through trained neural networks to accelerate MRI imaging.Therefore,based on deep learning technology,this paper conducted the following research for the above problems:Research 1: Generating Adversarial Networks Combining Attention Mechanism and Cycle Loss for Pelvic MRI SR ReconstructionCurrent single-contrast SR methods use structurally simple convolutional neural networks for MRI reconstruction,which can only extract limited information from MR images,resulting in poor reconstruction performance.Attention mechanisms and cycle loss have recently shown excellent performance in visual tasks.Attention mechanisms can help the model better understand the essential parts of the input data and focus on key regions,thereby improving the model’s performance and accuracy.Cycle loss is a loss function used in sequence generation tasks that can help the model better learn and maintain consistency between input and output sequences.Therefore,this study proposes a novel single-contrast SR method that uses a cycle loss and attention-driven generative adversarial network to reconstruct LR images into HR images.Experiments were conducted on pelvic data from healthy subjects and patients at 2× and 4× upscaling scales.Experimental results show that Research 1 uses the attention mechanism and recurrent loss to better recover the details in the original image.Research 2: Transformer-empowered multi-scale contextual matching and aggregation network for multi-contrast MRI SR reconstructionMRI can present images with the same anatomical structure but different contrast.Therefore,multi-contrast MRI SR reconstruction can utilize different but complementary information embedded in different imaging modes to reconstruct images with higher quality than single-contrast SR reconstruction methods.However,existing methods still have two drawbacks:(1)they ignore that multi-contrast features at different scales contain different anatomical details and lack effective mechanisms to match and fuse these features to obtain better reconstruction results;(2)they still have shortcomings in capturing long-range dependencies,which are crucial for regions with complicated anatomical structures.Therefore,this study proposes a Transformer-empowered multi-scale contextual matching and aggregation network to address these issues comprehensively.Firstly,the Transformer is trained to simulate the long-range dependencies between the reference and target images.Then,a novel multi-scale contextual matching method is designed to capture the corresponding feature information from different scales of reference features.In addition,a multi-scale aggregation module is introduced to gradually and interactively aggregate multiscale features to reconstruct the target image.Experiments show that Transformerempowered multi-scale matching and aggregation networks can model long-range relationships and reconstruct diagnosable MR images,which have great potential for application in clinical practice.Research 3: Synergizing wavelet and cross-attention Transformer for multi-contrast MRI super-resolutionCurrent multi-contrast MRI methods typically employ convolutional neural networks for feature extraction and fusion.However,existing models suffer from several drawbacks that hinder them from producing more satisfactory results.First,some high-frequency details in the image are lost during feature extraction,leading to blurry boundaries in the reconstructed image,which may impede subsequent diagnosis and treatment.Second,the limited receptive field makes it difficult for neural networks to capture long-range/non-local features.Third,these methods lack an effective multi-contrast feature fusion mechanism,which fails to utilize the complementary information in the reference image fully.Therefore,this study proposes a novel model that combines wavelet transform with cross-attention Transformer to address the above problems comprehensively.Firstly,a 2D discrete wavelet transform is used to obtain detail and approximation coefficients in the reference image.Then,a new residual cross-attention Swin Transformer is designed to extract and fuse multicontrast features to establish long-range dependencies between features and maximize the restoration of high-frequency information in the target contrast image.Additionally,a multiresidual fusion module is designed to merge the upsampled target image with highfrequency features from the original reference image,ensuring the restoration of detailed information.Experiments show that the combined wavelet and cross-attention Transformer can effectively reconstruct the high-frequency detail information in MR images.First,this master’s dissertation employs a GAN with attention and cyclic losses to enhance the quality of single-contrast MRI SR.Additionally,based on the characteristics of MRI,two Transformer-based multi-contrast MRI SR methods are designed to fully exploit the information in the reference contrast images for improving the reconstruction accuracy of the target image.Finally,the entire content is summarized,and future research directions are discussed. |
PDF Full Text Request