Classification Of Alzheimer Disease And Lesion Segmentation Combined With Semi Supervised Learning

Computer-based diagnosis of Alzheimer’s disease has developed greatly in recent years,but most of the existing research was carried out by means of using supervised learning.However,the labeling of medical images requires corresponding medical knowledge,leading to the scarcity of labeled medical image data.It in turn affects the progress of research in this field.Although there is less labeled data in this aspect,a large amount of unlabeled data is not used due to their lack of labels.Therefore,this thesis is going to investigate how to use those unlabeled data to improve the performance of Alzheimer’s disease lesion segmentation and classification.At the same time,the characteristics of medical images are different from those natural images.Therefore,how to establish a network model suitable for the extraction of complex medical image features remains a key issue.In response to the above problems,the specific work done in the current study is as follows:（1）Based on Unet and cross-consistency training,this thesis proposes a segmentation model for Alzheimer’s disease lesion segmentation.Some improvements have been made to Unet in this model.In feature extraction,the attention structure CBAM is used and in the skip connection process,SENet is employed.The main purpose is to make sure that the model can fully combine high-level semantic information with low-level semantic information to strengthen the semantic features of segmented parts and reduce the semantic features of non-segmented parts.In addition,the current study has also applied the cross-consistency training method to the model,so that the model can make full use of numerous unlabeled data to enhance the overall performance of the model.（2）Based on DenseNet and interpolation consistency training,this thesis puts forwards a classification model for classifying Alzheimer’s disease.In order to save the parameters of the model and improve the generalization ability of the model,the current study uses DenseNet as the backbone network.What is more,the method of interpolation consistency training is used in the model,so that the model can achieve better performance in training with only a few labeled data and a certain amount of unlabeled data.This thesis conducts experiments on both Alzheimer’s disease classification and lesion segmentation methods.In the segmentation task,the current study uses ACC and m IOU as indicators to evaluate the performance of the model in an objective way.The experimental results show that the segmentation method proposed in this thesis is optimal.In the classification task,the current study uses the evaluation index ACC to objectively evaluate the results of several two-class and three-class categories The experimental results indicate that the performance of the method proposed in this thesis is the optimal either in the two-class task or in the three-class task.Most importantly,either in the segmentation or the classification task,the methods used in the current study are based on semi-supervised learning,which can reduce the time of data labeling and make full use of unlabeled data to improve the performance of the internet-based model.