Medical MRI Image Processing Research Based On Augment Deep Learning

Image classification is an important exploration direction in image processing research.In medicine,the classification of medical images is of great significance for the study of intelligent diagnosis.In this paper,medical image classification is performed by deep convolution model,and data augmentation is realized by generating adversarial model,which can effectively identify different lesions.The work and writing of this article mainly include the following points:(1)For the uniqueness of medical images different from natural images,the advantages and disadvantages of classical deep convolution model are compared and analyzed.The DenseNet model is selected as the basic model of medical image classification,and the principle ?characteristics and advantage of DenseNet model are introduced in detail.(2)In view of the scarcity of medical image datasets,this paper uses data augmentation to augment the dataset,and designs a conditional gradient generation adversarial network CDCWGAN-GP model for data augmentation.Among them,the innovation of CDCWGAN-GP model mainly includes: 1.Using the Wasserstein distance with gradient penalty term instead of JS divergence in the optimization formula,effectively solving the problem of unstable GAN training,easy to generate gradient dissipation and Mode collapse.On this basis,the label y is added to the input of the network generator to guide the generator generation process;3.The discriminator is divided into a global discriminator and a local discriminator,and finally the outputs of the two discriminators are connected together for scoring,This allows the discriminator to fuse global features and local features.(3)This paper optimizes and improves the DenseNet model.There are three main improvements: 1.Reduce the redundant connections in the DenseNet model,analyze the necessity and practicability of reducing the connection in detail;2.Introduce group convolution in the dense block,so that Effectively reduce model parameters without adding additional parameters.3.According to the characteristics of the medical image dataset,appropriate model parameters are selected,including the growth rate K value,learning rate and optimizer.Finally,it is verified by experiments that the improved model effectively improves the computational efficiency and classification accuracy.