| Clear medical images are crucial for doctors to diagnose and treat patients quickly and accurately.However,due to limitations such as equipment,imaging principles,and patient safety considerations,obtaining ideal high-resolution medical images can be challenging.Super-resolution reconstruction is the process of restoring low-resolution images to highresolution ones.Deep learning-based SR methods can produce images with clear details and textures,which closely resemble real medical images and are easy to implement.Therefore,this paper focuses on deep learning image SR reconstruction techniques based on backprojection,attention mechanisms,feature fusion,conducts an analysis and research on these methods.The specific research contents are as follow.(1)To address the problem of low utilization of intermediate feature maps in current SR reconstruction networks and the insufficient use of information in the original LR images during the reconstruction process,a pyramid attention-based back projection super resolution reconstruction network is proposed.This network consists of an improved back-projection block,an efficient pyramid channel attention module,and a back-projection reconstruction module.By using the improved back-projection block to directly increase the SR image features during the deep feature extraction phase and feeding back the error of each scale to the network to correct the sampling results,the network performance is improved.In addition,the efficient pyramid channel attention module is used to fully extract multi-scale spatial information and important cross-dimensional features from the channel attention vectors.During the reconstruction phase,the back-projection reconstruction module is used to better utilize the information in the original LR images to produce better reconstruction results.The network achieves a PSNR of 28.98 d B at 4× scale factor on the LIDC dataset,which,when compared to other algorithmic networks,results in better quality SR images.(2)A high-efficiency back-projection-based super-resolution reconstruction network based on Shuffle Mixer is proposed to address the issue of model parameter redundancy in the super-resolution reconstruction process.The Shuffle Mixer feature fusion block is introduced to obtain both local and non-local information from the feature map,resulting in better feature representation.Channel splitting and channel transformation strategies are combined to reduce model parameters and computational costs.Furthermore,the network’s SR performance is further improved by using a convolutional attention module that integrates channel attention and spatial attention.Finally,a back-projection reconstruction module is used to obtain the SR image.The network has been experimentally validated to achieve a 1.384 M reduction in the number of parameters and a 0.86 d B improvement in the PSNR value compared to the back-projection super-resolution reconstruction network based on efficient pyramidal attention,resulting in better reconstruction results.(3)To further improve the generalization of the network model,a probabilistic degradation model is used.The probabilistic degradation model learns how to generate the degradation distribution of real images,thereby synthesizing a large number of training samples that conform to the distribution of real medical images.By integrating the probabilistic degradation model with the proposed efficient back-projection super-resolution reconstruction network based on Shuffle Mixer,it has been experimentally demonstrated that the network model can be better applied to real-world environments. |
