A Suspected AD Brain MRI Image Classification Algorithm Based On Multi-core Learning

Senile dementia is a neurodegenerative disease associated with various etiological factors.The typical clinical symptoms of senile dementia are mental and behavioral disorders caused by degradation and damage of neurons.Examples of such disorders include the deterioration of memory and other advanced cognitive functions,the loss of daily living ability and the decline of abstract thinking ability.Alzheimer disease(AD)is one of the most common senile dementias which bring severe stress and serious harm to patients and their families.It also imposes substantial financial burden on the whole society.Therefore,prevention,diagnosis and treatment of AD are extremely urgent and increasingly important.Magnetic resonance imaging(MRI)image analysis is one of the most active research areas in image processing theory and application.It is also an important method for clinical diagnosis for patients with AD.Aiming at existing problems in current AD-related MRI image classification,in this thesis,we focus on algorithms for automatic classification of AD-related MRI images based on vector multi-kernel learning and tensor multi-kernel learning.At the same time,based on the basic theory and the precondition of kernel function operation in the same space,this thesis provides an in-depth analysis on how to effectively combine kernel functions respectively associated with spatial and anatomical structure information in a MRI image,and how to reduce the complexity involving in spatial kernel function computation.The kernel functions containing spatial and anatomical information based on the above theory are applied to the classification of AD-related MRI image data.The main contents and innovations are as follows:Firstly,the existing classification algorithm of AD-related MRI data based on traditional vector single-kernel learning is studied and the advantages and disadvantages of the classical Cuingnet framework are also analyzed.In the related work,Cuingnet proposed a general classification framework to introduce spatial and anatomical structure information in classical single-kernel support vector machine(SVM)optimization scheme for brain image analysis,which has achieved good classification accuracy.However,this framework has two main drawbacks.First,it involves spatial and anatomical regularization and in order to satisfy the optimization conditions required in the single kernel case,it is pre-assumed that the spatial regularization parameter value is equal to the anatomical one.Second,in the framework it has to convert a 3D discrete brain image,which is generally and naturally represented by a higher-order tensor,to a one-dimensional vector in order to meet the input requirements.In this manner,the natural structure in the original data is destroyed,and what is worth,generally it produces a very high-dimensional vector so that the famous problem of curse-of dimensionality in machine learning will become more serious,and at the same time,a huge adjacency matrix is unavoidably adopted during the construction of spatial kernel,to determine the adjacency relation between each pair of voxels,which leads to very high time and spatial complexity.To overcome the first drawback in the Cuignet framework,i.e.,spatial and anatomical regularization parameters are pre-assumed to be equal,in this thesis we propose an improved method named Spatial-Anatomical-MKL(multiple kernel kearning)to combine the sequential minimal optimization(SMO)algorithm with MKL to solve the optimization problem of regularization parameter estimation.We first prove that the spatial and anatomical kernel functions in the Cuingnet framework can satisfy the preconditions of the Kloft model and the SMO-MKL algorithm,so that the spatial and anatomical kernel functions can be linearly combined and the weight coefficients of these kernel functions can be properly determined.Therefore,the Spatial-Anatomical-MKL method effectively solves the first problem in the Cuingnet framework and it also extends the algorithm for classification of AD-related MRI data from vector single-kernel learning to vector multi-kernel learning.However,the second problem in the Cuingnet framework,that is,a huge adjacency matrix is unavoidably adopted to determine the adjacency relation between each pair of voxels,still remains unsolved in the Spatial-Anatomical-MKL method and thus the computational complexity is still high.To overcome the second drawback of high computational complexity in spatial kernel calculation in the Cuignet framework,this thesis proposes an improved method named Spatial-Prior-in-STM based on support tensor machine(STM).First,we perform a more detailed analysis of alternating iterative algorithm of classical rank-1 STM and then proposed an improved version wherein the iterative process is divided into two steps so that the canonical decomposition / parallel factors(CP decomposition)and spatial structure information of all frontal slices of an MRI image is included in the classic rank-1 STM model,in a manner that many smaller adjacency matrices of frontal slices are adopted to retain the spatial structure information and to reduce the space and time overhead.However,due to the limitation of the alternating iterative algorithm,the Spatial-Prior-in-STM method only contains the spatial structure information of all frontal slices of MRI data and cannot include the spatial structure information of any horizontal and lateral slices and the anatomical structures information of all kind of slices of MRI data simultaneously.To tackle the above drawbacks of Spatial-Anatomical-MKL and Spatial-Prior-in-STM method and provide a comprehensive solution to the two problems in the Cuingnet framework,two new tensorial kernels named Zero-Extended-Kernel and Frontal-Horizontal-Kernel are constructed so that the precondition of the Kloft model and the SMO-MKL algorithm is satisfied,and thus these kernels can be linearly combined with proper weight coefficients.Furthermore,the two new kernel functions use many smaller adjacency matrices of all kind of slices to retain the spatial and anatomical structure information of an MRI image.In this manner,the high computational complexity problem in the Cuingnet framework is efficiently solved.Theoretical analysis and experimental results show that adoption of the above two kernel functions not only greatly improve the computational speed of spatial kernel calculation,but also maintain high classification accuracy rates.Therefore,Zero-Extended-Kernel and Frontal-Horizontal-Kernel tensorial kernel extend the algorithm for classification of AD-related MRI data from vector single-kernel learning and tensor single-kernel learning to tensor multi-kernel learning.