| Image classification task is an image processing method to achieve image classification and recognition by extracting feature information of images.With the development of computer technology and the improvement of medical hardware performance,a large amount of medical image data has emerged to improve and strengthen the computer-aided diagnosis technology.As a basic and key technology in the field of computer vision,many classification network models are used in modern medical-aided diagnosis to largely reduce the influence of external environment and operators on medical image acquisition.However,for Parkinson’s disease,there are still some limitations in the current research.The small macroscopic variance of medical images requires reducing the loss of key pathogenic information when extracting features from the network.Secondly,the specialized and complex annotation process of medical images with limited labeled data samples and the demand of data for deep learning is large,which limits the development of deep learning in the field of Parkinson’s disease.In response to the above problems,this paper proposes two models for the classification of MRI images of Parkinson’s disease brain,as follows:(1)A deep convolutional network based on attention mechanism is proposed for medical image classification.The classical deep convolutional network model Dense Net is chosen as the basis,and a global hybrid attention module is proposed to be introduced into the network structure so that the model acquires global information and channel information of the feature map to solve the problem of critical information loss in capturing brain tissue location and content information in deep convolutional networks due to the limitation of local sensory fields;pruning the connection methods of the network to remove unnecessary ones and retain The more useful and fast connections are removed to alleviate the overfitting caused by the image characteristics and the small size of the data set.The improved Dense Net network model achieved 85.44%classification accuracy on the PPMI dataset,which is 8.99% higher than the accuracy before the improvement.(2)From the perspective of the practical application scenario of medical images,a large number of unlabeled data samples are utilized and a deep convolutional adversarial network based on coding transformation and streamwise regularization is proposed to address the problem that the random noise is used as the starting point of the generator to generate samples in the traditional generative adversarial network,which cannot fully learn the feature distribution of the real samples and makes the offset between the discriminator extracted data features and the real data features.network model based on coding transformation and stream regularization.The network structure of the generative adversarial network is improved and the encoder network is introduced to map the original data samples into coded hidden variables with the information of the original sample image features,and the noise hidden variables are obtained by the reparameterization method instead of random noise as the starting point of the generator;the stream shape regularization is introduced in the discriminator to keep the perturbation constant in the direction of the data stream shape and to label the similarity of the closer points to Improving the generalization ability of the model.The improved model achieves 87.18% classification accuracy on the PPMI dataset under the semi-supervised classification task. |
