Super-resolution Reconstruction Of Medical Images Based On Deep Learning

Magnetic resonance imaging is a safe technique that uses powerful magnets,radio waves,and computers to create images.MRI can image most parts of the body in any direction to obtain information and provides this information in high-quality images.When using the magnetic resonance image for non-invasive medical treatment,the clarity of the image affects the doctor's judgment.High-resolution images can provide doctors with more abundant pathological information and improve the credibility of the diagnosis.In magnetic resonance imaging,acquiring a high-resolution image requires a longer scan time and a higher signal-to-noise ratio.However,it is difficult for the patient to remain still for a long time,and the high signal-to-noise ratio requires more sophisticated instruments,which increases the imaging cost.In order to shorten the scanning time,the commonly used method is to increase the scanning layer thickness,but this will lead to a decline in resolution,and ultimately will limit the processing,analysis and diagnosis of the disease in later stages.Super-resolution reconstruction is a technique for reconstructing high-resolution images using single or multiple low-resolution images.This technique is widely used in medical,remote sensing,security,and other fields.Obtaining high-resolution magnetic resonance images by super-resolution reconstruction algorithm is a feasible solution because it is inexpensive and does not require the patient to scan for a long time.In recent years,with the development of machine learning and deep learning technologies,learning-based super-resolution reconstruction algorithms have been well studied.Deep network structures can extract features from the data to build more and more abstract representations,replacing traditional methods of manually extracting features and manually creating algorithms.In this paper,an image imaging model is established to study the use of convolutional neural networks for super-resolution reconstruction,and this method is applied to brain magnetic resonance images.The main contents of this study and innovative research results are as follows:(1)Full convolutional neural network has achieved very good results in the field of image segmentation.The full convolutional network can accept any size input and recover the image size through the deconvolution layer.This paper uses the characteristics of convolutional neural networks to design a feed forward full convolutional neural network to perform super resolution reconstruction of two-dimensional magnetic resonance images.In order to improve the visual sense of reconstruction,the output of the network is re-inputted into the VGG-16 network,and the lower-level features are extracted to calculate the perceived loss error.The model achieves 4x magnification of the image inside the model while avoiding the problem of using the deconvolution layer prone to artifacts.The experimental analysis compares several different super-resolution reconstruction algorithms.The reconstructed results of the designed network are more in line with visual perception.The two indicators of peak signal-to-noise ratio and structural similarity are increased by about 1% and 2% respectively.(2)The three-dimensional convolutional neural network is mainly applied in the fields of behavior identification and scene recognition.The three-dimensional convolution kernel can be used to extract the spatial features between the sequences before and after the image.The magnetic resonance image itself has a three-dimensional spatial characteristic,so a convolutional neural network containing a three-dimensional convolution is designed to obtain spatial information of the magnetic resonance image sequence to guide super-resolution reconstruction.The network mainly consists of three parts: block extraction and representation,nonlinear mapping,and reconstruction.Compared with the proposed 2-D magnetic resonance image super-resolution reconstruction method based on perceptual loss,the peak-to-noise ratio and structural similarity were improved by about 4% and 3% respectively.