The Research Of Alzheimer’s Disease Classification Based On Brain Image Features

With the increasing aging of the global population,Alzheimer’s disease(AD)has become an irreversible neurodegenerative disease commonly in elderly adults,which imposes a heavy financial burden on the families and society.Mild cognitive impairment(MCI)is an intermediate state between AD and normal aging/cognitively normal older individuals(CN).Studies have found that timely treatment of patients in theMCI stage can minimize or prevent their progression to AD.Therefore,the development of methods to accurately distinguish AD,MCI,and CN has important clinical and social significance.With the continuous development of medical imaging technology,brain images such as structural magnetic resonance imaging(s MRI),diffusion tensor imaging(DTI),functional magnetic resonance imaging(f MRI),and positron emission computed tomography(PET)have been widely used in the clinical diagnosis and detection of AD.However,the method of diagnosing diseases by directly viewing brain images may be time-consuming and highly subjective for clinicians.In addition,some features of brain images are not obvious to human vision that clinicians cannot directly observe them.With the advent of the internet era,various image processing and analysis techniques based on natural images have been developed rapidly,and researchers have begun to apply them to feature extraction of brain images.However,due to the unique complexity of brain images,most of the image processing and analysis methods based on natural images are often difficult to achieve satisfactory results in brain images.In this study,we focus on feature extraction techniques for s MRI images and PET images,then combine traditional machine learning and deep learning techniques to construct a classification model framework and apply it to binary classification tasks as well as multi-classification tasks for AD.The specific research contents and their results are as follows:(1)The features extracted using voxel-based methods in s MRI images usually have high dimensionality,which leads to potential over-fitting problems in classification tasks.To address this problem,a new feature extraction method of equal-distant rings shape context(EDRSC)is proposed for feature extraction of s MRI images and applied to the classification study of AD.Firstly,the region of interest was extracted from s MRI images.Secondly,the information entropy of all slices of the region of interest was calculated to select the required 2D image.Next,a visual attention model was used to detect the saliency map of the2 D image,and an edge detection operator was used to detect the shape contour of the saliency map.Then,the shape contour of the saliency map was segmented by equal-distance ring to extract its shape context features.Finally,support vector machine(SVM)classifier was used to construct the classification model.Classification experiments were performed using s MRI images of AD subjects,MCI subjects,and CN subjects on the Alzheimer’s Disease Neuroimaging Initiative(ADNI)dataset.The experiment results show that in binary classification tasks,EDRSC features of s MRI images can achieve higher classification accuracy compared with existing features such as gray matter volume,cortex thickness and complexity.Among them,the classification accuracy rates in AD and CN classification and AD andMCI classification were all greater than 93%.(2)A non-negative matrix factorization-tensor decomposition network(NMF-TDNet)is designed to address the data dependency and "dimensional disaster" problems of lightweight deep learning network-principal component analysis network(PCANet).It is also applied to the feature extraction of 3D s MRI images,and then combined with SVM classifier,an AD classification method based on lightweight deep learning network features of s MRI images is proposed.Using 3D s MRI images in the ADNI dataset as the experimental subjects,the experimental results show that NMF-TDNet method can achieve data dimensionality reduction(the dimensionality of the extracted features numbers only a few hundred dimensions,far less than the hundreds of thousands required by PCANet method),thus reducing the storage cost and improving the computational efficiency.At the same time,the experiment results of binary classification tasks show that the features extracted by the NMF-TDNet method achieves better classification performance than the features extracted by the PCANet method,and its classification accuracy was improved by at least15.03%.(3)Considering that different modalities images can provide different information for the classification and diagnosis of AD to improve the classification accuracy of AD,a method combining three-dimensional discrete wavelet transform(3D-DWT)and three-dimensional invariant moments(3D-MIs)is proposed for feature extraction of 3D s MRI images as well as 3D PET images,and these features are used as input to a neural network to construct a classification model.Firstly,the automated anatomic labeling(AAL)brain atlas was used to identify brain regions of interest in 3D s MRI images and 3D PET images,respectively.Secondly,each brain region of interest was decomposed using 3D-DWT to obtain its approximate components,and then the shape features and spatial features of the approximate components were quantified using the 3D-MIs method.Finally,a stacked autoencoder(SAE)network containing two hidden layers and a softmax classifier was used to construct the classification model.Eight groups of binary classification task experiments were conducted on the ADNI dataset,and the experimental results show that the proposed method can well distinguish AD fromMCI and CN,and the classification accuracy was 84.29% and 96.92%,respectively.(4)To address the problem that existing features fusion methods for different modalities images have the problem of extracting features independently from a single modality image and then simply concatenating them into a long vector for classification without properly considering the correlation between different modalities images,a method based on multi-modalities images brain networks and quaternion convolutional neural networks(QCNN)is proposed for classification and diagnosis of AD.Specifically,firstly,according to the AAL brain atlas,different modalities images were divided into brain regions to construct different brain networks,then the different brain networks were synthesized into a quaternion brain network matrix,and finally the quaternion brain network matrix was input into the QCNN for training to classify AD.The experimental results on the ADNI dataset show that the proposed method can better integrate the information between the different modalities images by combining the brain networks of different modalities images into a quaternion matrix,and achieve better classification results on four groups of binary classification tasks and two groups of multi-classification tasks.Especially in the pMCI and sMCI classification tasks,the classification accuracy was 89.74%.