Study On The Classification Of Alzheimer's Disease Based On Densely Connected Convolutional Networks

Alzheimer's disease(AD)is an irreversible neurode generative disease of the brain.Mild cognitive impairment(MCI)is a symptom between AD and cognitively normal(CN).As the early stage of AD,it is very important to prevent and intervene in the development of the disease.The classification of Alzheimer's disease based on DenseNet is studied and analyzed.An early diagnosis model of AD and a prediction model for the conversion of MCI to AD are proposed.The active learning algorithm is also studied and a multi-strategy batch mode active learning algorithm is proposed.The main research contents of this thesis are as follows:An early diagnosis model of AD based on multi-scale features and sequence learning is proposed.Aiming at the problem that 2D-CNN cannot effectively obtain the sequence information in MRI,and considering the small size of MRI dataset,a 3D-LDenseNet with less parameters is given as the basic model.Then,for the problem that 3D-LDenseNet can only extract image features in local receptive fields and local slice sequences,the dilated convolutional and the convolutional long and short-term memory are introduced into the 3D-LDenseNet to enhance the multi-scale feature extraction ability and sequence learning ability of 3D-LDenseNet for MRI.A conversion prediction model of MCI to AD based on attention mechanism is proposed.Aiming at the problem that the changes in lesion areas of progressive MCI(pMCI)and stable MCI(sMCI)are slightly different,a spatial attention mechanism and a channel attention mechanism are introduced into the early diagnosis model of AD to extract the salient feature of MRI.Then,the model can acuurately locate the salient regions that are help ful for classification during the training process,so as to enhance the ability of model to distinguish sMCI and pMCI.A multi-strategy batch mode active learning algorithm is proposed for Alzheimer's disease classification.Aiming at the problem that information redundancy in batch samples selected by batch mode active learning,on the basis of using the uncertainty strategy,the convolutional auto-encoders and the clustering algorithm are introduced to measure the representativeness of samples,and the distance measurement is introduced to measure the similarity between unlabeled samples and labeled samples.By comprehensively measuring the uncertainty,representativeness and similarity of the unlabeled samples,the representativeness of selected samples and the diversity of labeled samples are improved,and the redundant information between selected samples is relieved to a certain extent,so that as few labeled samples as possible can be used to achieve the expected performance of the model.