Research On MRI Reconstruction Based On Generative Adversarial Network

Nuclear magnetic resonance technology has been widely used in the current production and life.It is very popular in the fields of material analysis,medicine and so on,especially in the medical imaging diagnosis.However,in the process of MRI reconstruction,there are some problems,such as slow reconstruction speed and fuzzy artifacts.On the other hand,with the continuous exploration of deep learning neural network and other academic fields,a more advanced generative adversarial network is proposed.Based on the theory of generative adversarial network,this thesis studies the reconstruction algorithm of MRI.Firstly,this thesis proposes a reconstruction algorithm of MRI based on conditional generative adversarial network.In terms of composition,the algorithm uses the Unet structure as the generator network and the convolution neural network as the discriminator,and combines the pixel domain mean square error loss,frequency domain mean square error loss,perception loss and resistance loss into the loss function.Through the contrast experiment,it is verified that the algorithm has a significant improvement in image reconstruction speed compared with the traditional reconstruction algorithm,and the information quality of the reconstructed image is improved by more than 3%.In addition,it is also verified that the frequency domain mean square error loss is effective for improving the information quality and the perceptual loss is effective for improving the perceptual effect.Through the contrast experiment,it is verified that the algorithm has a significant improvement in image reconstruction speed compared with the traditional reconstruction algorithm,and the information quality of the reconstructed image is improved by more than 3%.In addition,it is also verified that the frequency domain mean square error loss is effective for improving the information quality and the perceptual loss is effective for improving the perceptual effect.But at the same time,we also find that the algorithm has no advantage in the perception effect and the reconstruction result has the problem of fuzzy sharpening.Then,in order to solve the problem of perceptual vision,this thesis proposes an MRI reconstruction algorithm based on residual dense connection and relative generative adversarial.In this algorithm,the residual dense connection blocks and the structure of the convolution layer are used to improve the generator network.The idea of relative generative adversarial network is used to improve the discriminator,and then the calculation method of perceptual loss in the loss function is changed to improve the perceptual effect of the algorithm reconstruction.Compared with the traditional method,the improvement of the reconstruction speed and the information quality of the reconstructed image is similar to the previous algorithm,but the perceptual effect is improved by more than 8%,and the reconstruction result has better visual effect.Finally,it should be noted that the reconstruction information quality of this method is about 0.13%lower than that of the previous algorithm.By sacrificing a small amount of information quality,the perception effect is greatly improved,and a more balanced state is achieved between the two indicators.