Early Diagnosis Of Alzheimer’s Disease Based On Deep Learning And Multi-scale Information Fusion

Alzheimer’s disease(AD)is a neurodegenerative disease and the most common type of dementia in the elderly.Its clinical symptoms are progressive memory loss and cognitive dysfunction,and because of its irreversibility and the enormous social burden it causes,an accurate diagnosis will help physicians intervene in a timely manner to treat the disease,and is expected to help patients and their families reduce long-term care costs.With the development and breakthrough of medical imaging and deep learning technology,in response to the problem of low accuracy of early diagnosis of Alzheimer’s disease,the use of deep learning algorithms to build automatic diagnosis models to replace the subjective judgment of doctors and accurately predict AD and MCI patients can help reduce the burden of society and families,which is an important initiative to promote the application of artificial intelligence in the medical system and can effectively solve the medical resources problems such as imbalance between supply and demand.However,there are many limitations in existing studies,such as the models do not consider the complementarity between multiple types of features,traditional neural network models cannot efficiently extract pathological features in the brain,and ignore the spatial topological properties in the brain structure.To address the existing problems,this thesis investigates the Alzheimer’s disease diagnosis method based on deep learning and multi-scale information fusion,and the main research includes the following two aspects.(1)This thesis proposes a DTI and multiscale information fusion-based method for early AD aided diagnosis.To address the problems of insufficient single feature characterization ability and low accuracy for disease identification,this thesis proposes the use of three feature linear fusion strategies to extract multidimensional feature vectors from DTI images of AD,MCI and NC subjects.Firstly,we extracted the white matter fiber bundle skeleton using the spatial statistical method of fiber bundle tracing,and extracted the voxel-based anisotropy score and mean dispersion rate features from the skeleton;secondly,we constructed a special 3D convolutional neural network model to extract 3D deep features directly from the whole brain;again,we constructed the brain structure network of each subject based on the fiber bundle connectivity relationships between different brain regions to extract brain connectivity features.Finally,the voxel-based features,connectivity-based features and deep features were linearly fused,and after feature selection by Relief F algorithm,was used to classify AD,MCI and NC using support vector machine.The experimental results show that our proposed method significantly improves the classification performance and can represent the sample features more efficiently compared to existing methods.(2)This thesis proposes a method of AD early aid diagnosis based on DTI and graph convolutional neural network.To address the problems of poor characterization of brain connectivity features,underutilization of spatial topological structure information,and low accuracy rate.We propose a new graph convolutional network model for learning discriminative features from structural brain network data.First,we extract spatial information of different brain regions from DTI data and model the inter-regional topology to construct the brain structure network.Then,we designed convolution filters from edge-to-edge,edge-to-node,and node-tograph for the characteristics of the DTI brain structure network,as a way to construct a graph convolutional network model and classify AD,MCI,and NC.The experimental results show that our proposed graph convolutional neural network classification model can provide accurate diagnosis of diseases,and can provide more accurate disease diagnosis results compared with traditional methods.