Research On Generative Adversarial Network Algorithm Based On MRI Data

With the development of deep learning,many network models have been applied to modern medical auxiliary diagnosis.In recent years,generative adversarial networks have received wide attention in the field of medical imaging.With the help of deep neural network structure and the unique two-player game training mechanism of generative adversarial networks,they are applied to tasks such as synthesis,segmentation,and reconstruction.The generative adversarial network model must include adiscriminator and a generator,the role of the discriminator is to achieve dual classifications,so it is actually a classifier by nature.The generator can be used to generate simulated samples with the same distribution as the original image,which is equivalent to adding a large number of training set samples.However,the current research still has some limitations.First of all,publicly avail-able high-quality data sets for Parkinson’s disease are currently scarce,and deep learning has a very large demand for data,which undoubtedly limits the development of deep learning in the field of Parkinson’s disease.Secondly,because the generative adversarial network is mainly used for image generation currently,its confrontation learning method causes the discriminator to be unable to generate high-quality simulation samples while achieving high classification accuracy.Thirdly,the difference between medical images and natural images makes most models unable to achieve as good results when processing medical images as when processing natural images.To solve the above problems,this paper proposed two models based on generative adversitynetwork for the generation and classification of MRI images of Parkinson’s disease.The specific research contents are as follows:(1)For the difference between medical images and natural images,an attention-based mechanism and progressive network are proposed(Attention Progressive GAN,APGAN).By introducing progressive network into the network structure,the model can better learn the feature distribution of samples with different resolutions.Attention mechanism is introduced to avoid the problem of low global consistency of generated images due to the convolution kernel size limitation of local receptive fields.Smooth transition technique and other training techniques are used to enhance training stability.FID and PSNR indicators were used to measure the quality of the generated images,so that the generated images could better fit the real sample distribution.(2)Aiming at the problem that the generative adversarial network is difficult to balance the generated image quality and classification accuracy,a kind of triple-progressive GAN(TPGAN)network based on three-person game and manifold regularization is proposed.A classifier was added to turn the two-player game into a three-player game,making the training process more stable.Then,a progressive network is used to replace the original network structure,which makes the images generated by the generator more difficult to distinguish when processing high-pixel and large-size images.Manifold regularization is introduced into the classifier to encourage the classifier to keep the local perturbations of generator parameters unchanged,and to label the points in the data manifold similarly to improve the generalization ability of the model.Experimental results on PPMI data set show that both APGAN and TPGAN have better performance in image generation and image classification than current baseline.