Research On Brain Magnetic Resonance Imaging Statistical Classification Methods For Diagnosis Of Schizophrenia

Schizophrenia is a serious mental disability.It is generally believed that accurate diagnosis at the early stage of schizophrenia plays a very important role in controlling the progress of schizophrenia effectively.However,at the early stage of schizophrenia,because there is no obvious behavioral abnormality,diagnosis becomes very difficult.In this context,Magnetic Resonance Imaging(MRI)technology,as an important medical imaging technology,becomes popular in early diagnosis of schizophrenia.How to use statistical classification methods to analyze MRI brain images and realize automatic diagnosis of schizophrenia is always an important research topic in related research fields.However,due to the complexity of schizophrenia itself,the high dimensionality and the small sample size of MRI data,the effectiveness of currently automatic diagnosis methods for schizophrenia is not satisfactory.Under the support of the key scientific research project from the Education Department of Hunan Province "research on machine learning methods for spatially structural analysis of brain imaging data and its realization using heterogeneous computing technique",in this paper we mainly do research on effective machine learning methods for solving heteroscedasticity,high dimensionality and the small sample size problems and apply them to automatic diagnosis of schizophrenia.Our main works are summarized as follows:A regularized weighted least squares classifier based on optimization technology is proposed in this paper.2 classes of regularization functions satisfying 2 basic properties are constructed.The basic properties of these 2 regularized functions are theoretically analyzed,and corresponding iterative optimization algorithms are designed.Comparative classification experiments on artificial data sets and UCI data sets are carried out.The experimental results show that the proposed classification method is very robust and is more suitable for the classification of heteroscedasticity than other benchmark methods.Finally,we do classification experiments on COBRE dataset(a freely brain MRI dataset),extract single layered functional connectivity network,achieve better classification results and improve the robustness and stability of classification effectively.A deeply nonnegative supervised autoencoder neural networks is designed in this paper.This proposed model is based on stacked deep autoencoder neural networks and has 3 new features for improving the generalization performance of deep neural networks in the case of small sample size.The first feature is introducing feature selection using L1 norm minimization at the beginning stage.The second feature is learning supervised features by introducing a new regularization item of correlation coefficients.The last feature is controlling network capacity by introducing nonnegative constraints of weights.To solve backpropagation of the new loss function with constrained item,we further design the updated rule of weight which meets the requirements and prove the rule is feasible in theory.Finally,the proposed deep autoencoder neural network model is applied to the COBRE schizophrenia dataset.The classification effect is further improved and multi-layered brain functional connectivity networks are extracted successfully.An orthogonal features generation method for convolutional neural networks is proposed in this paper.By introducing heterogeneous orthogonal degree and similar correlation degree into optimization objectives,using the proposed orthogonal features generation method can help to improve the discriminative features abstracting power of neural networks,enhance the sparsity of middle layer outputs.All of these play a significant role on controlling the network capacity,improve the generalization and the power of local discriminative region location.To improve the computation efficiency of discriminative orthogonal features generation,a random 2-class identification method is proposed.an adaptive regularization coefficient adjusting method is further proposed to control the sparsity of network outputs and approximate the preestablished threshold.A comparative experiment for small sample size problem on MNIST handwritten numeric dataset is carried out.The results show that the performance of the proposed discriminant orthogonal feature generation is good in the case of small sample size.Finally,we carry out classification and comparative experiments on COBRE schizophrenia dataset using structural magnetic resonance brain imaging data and achieve the better classification effects.By using deconvolution visualization operation,we also successfully locate the local discriminating brain area.