Research On Super-resolution Algorithm Of MR Images Based On Deep Learning

Magnetic Resonance Imaging is one of the most important imaging modalities in medical imaging.It has no radiation damage and can acquire both three-dimensional and cross-section images of the body.Spatial resolution is one of the significant param-eter of Magnetic Resonance(MR)images.High-resolution(HR)MR images have rich details and accurate information about lesions,which play an important role in clini-cal diagnosis.However,the resolution of MR images is restricted by various factors,higher resolution images could lead to signal-to-noise ratio(SNR)decreasing and pro-long scanning time.Super-resolution(SR)technology is the most effective method to improve the resolution of MR images.It can complete the conversion of low-resolution to high-resolution images without changing the hardware conditions,so it has received widespread attention.This paper mainly conducts a study on SR algorithms of MR images based on deep learning methods.First,this study proposes a new type of network structure(Hybrid-Net)to improve the resolution of MR images.Based on the dense blocks,the network introduces a multi-path structure to extract rich MR image features.Lots of experiments are conducted on multiple MR datasets.The experimental results show that the algo-rithm not only reconstructs high-quality MR images,but also reduces the computational complexity of the network and improves the efficiency of network reconstruction.In addition,an unseen MR dataset is used to test the network.The results have proved that the HybridNet has good robustness and generalization ability,and can reconstruct clear HR images.In practical applications,people often need to enlarge images by different factors.The network which with a single amplification factor is difficult to meet its needs.If the network is trained separately for each amplification factor,it not only wastes comput-ing resources,but also complicates the application process.In order to meet the above needs,this paper builds a SR algorithm model(MFSR)for multiple magnification fac-tors,and adopts two MR datasets to train and test the model.From the experimental results,it can be seen that MFSR can reconstruct clear HR images under multiple mag-nification factors.Even for pathological MR images,it also can accurately enlarge the lesion area without introducing obvious interference information and visual artifacts.Overall,this paper adopts deep learning methods to build two neural networks,completed the MR SR tasks of single and multiple magnification factors.The algo-rithms proposed in this paper can reconstruct clear MR images efficiently.