| As one of the important reference bases in clinical diagnosis,medical images can provide information about human organs and help doctors determine the location and shape of diseased tissues.However,medical images are subject to varying degrees of degradation during imaging,storage and transmission,which will directly affect the physician’s analysis of the patient’s true condition.Super-resolution technology can enhance the texture details and boundary structures of low-resolution images to improve the clarity and resolution of images.At present,most of the super-resolution reconstruction algorithms for medical images are scene migration and improvement of natural image reconstruction algorithms,but medical images have the visual characteristics of complex structure,inconspicuous details and more invalid information,which make the reconstruction effect slightly insufficient.To address the above problems,thesis proposes a super-resolution reconstruction network DHAM-GAN for medical images based on generative adversarial networks,which uses a dynamic hybrid attention DHAM module in the generator part to learn the feature mapping between low-resolution to high-resolution,and drives the generator to produce more realistic high-resolution images through relative discriminators.The DHAM module has a dual-layer of attention and is able to dynamically adjust the features after adaptive weighting of the mixed attention branches,improving the network’s ability to focus on important information.In addition,DHAM-GAN uses a new total variational loss to minimize the gradient difference between the reconstructed image and the original image at neighboring pixel points,so that the reconstructed image produces some sharpness in the detail region.Experiments show that DHAM-GAN achieves better results on the reconstruction of CT images,but poor visual perception on MRI reconstructed images.In order to further improve the reconstruction quality of medical images,thesis proposes an RMSA module based on the self-attention principle and improved residual structure.This module calculates self-attention weights for segmented sub-features using different scale windows,adding more self-similarity information while controlling the computational effort.Another network RMSA-GAN for super-resolution reconstruction of medical images is designed based on the RMSA module,which can efficiently capture local structures and longterm dependencies.Meanwhile,to be able to use multi-level features for reconstruction,the network performs feature fusion on the output of the RMSA module at all levels.Besides,the use of gradient variance loss in the loss function of RMSA-GAN forces the model to generate high gradient variance on the reconstructed images,improving the network’s ability to reconstruct high-frequency details and sharp contours of medical images.Thesis,we propose two super-resolution reconstruction algorithms for medical images based on deep learning techniques,which can effectively recover detailed textures in images and address the need for high-quality images in clinical settings at the algorithmic level.Therefore,this study can provide some technical support for the development of medical imaging technology and clinical medicine. |
