Generative Adversarial Networks For The Classification Of Liver Lesion

Image classification has always played an important role in the field of computer vision.With the continuous development of various image acquisition technologies,more and more images have entered people’s lives.The existing image classification technology can classify images efficiently and accurately.It is well known that the training process of the classifier requires a large number of samples.However,in the field of medical images,it is found that there are still problems such as limited data sets and limited number of images with annotated samples.People hope to have more images to help people train classifiers with higher classification accuracy.Since the launch of the Generative Adversarial Nets,it has received a wide range of attention.The emergence of the Generative Adversarial Nets is of great significance for the development of the generative model.The most direct application of GAN is to generate data,in other words,to use GAN to generate images,voices,and other data.As the most successful model in the field of image processing,Convolutional Neural Network have been widely applied in many fields.As early as the end of the20 th century,Convolutional Neural Network have been applied to medical image analysis,and later affected by factors such as gradient disappearance.So that research on Convolutional Neural Network has entered a trough.Until people have found a way to train deep neural networks,so that Convolutional Neural Network have been reapplied to the field of medical imaging,how to combine CNN and GAN has become the direction of many scholars,and DCGAN is in this respect.One of the best try.However,the convolutional network of the DCGAN model is limited bythe local receptive field,and it is unable to generate a wide range of related regions,which ultimately leads to the inability of the generated images to achieve the desired results.In response to the above problems,this paper proposes an improved DCGAN model.The model introduces the self-attention mechanism into DCGAN,and replaces the convolutional feature map generated by convolution in the original DCGAN with a self-attention feature map.The problem that the model is limited by the local receptive field and the convolutional network cannot generate a large range of related regions is solved.This paper is mainly composed of the following parts:(1)this paper introduces the basic knowledge of the Generative Adversarial Nets,including what is the Generative Adversarial Nets,several classical variants of the Generative Adversarial Nets and the key points of related knowledge used in this paper.It also introduces the research status of the Generative Adversarial Nets and the application of the Generative Adversarial Nets in medical image field.(2)Introduced the advantages of the Generative Adversarial Nets compared with traditional data enhancement methods,and focused on the DCGAN model,which combines convolution neural networks with the Generative Adversarial Nets,can effectively improve the quality of generated samples and convergence speed.(3)Because of the limitation of data set size,the original data set is processed by the method of three fold cross validation,which ensures that every data in the data set has the same probability to be used for training and testing,and improves the generalization ability of the model to a certain extent.Because DCGAN model is limited by local receptive field,the convolution network can not generate a large range of related areas.This paper improves the DCGAN model.The self-attention mechanism is introduced into the model,so that the improved model can generate higher quality pictures,thus improving the accuracy of classifier classification.(4)In order to verify the actual performance of the DCGAN model based on the self-attention mechanism in the classification of liver lesion,this paper compares the DCGAN model based on the self-attention mechanism with the classical data enhancement method and the original DCGAN model,using the confusion matrix,Sensitivity were used to evaluate the performance of the classifier.The final experimental results show that the DCGAN model based on the self-attention mechanism has better performance in the classification of liver lesion.