| The brain is one of the most important organs of human beings,and one of the important objects of medical and neurophysiological research.Although brain science has made great progress through years of research,the brain diseases that have a major impact on patient,their families and society,have never been cured.It is generally believed that early detection and early intervention are of great significance for the treatment of brain diseases.When patients with brain diseases do not show obvious symptoms in the early stage,Magnetic Resonance Imaging(MRI)can provide imaging support for the early diagnosis of brain diseases.Statistical machine learning is the main method of brain image analysis that is indispensable at this stage.With the rise of deep learning,how to use deep learning techniques to realize the classification of MRI brain images to assist the automatic diagnosis of brain diseases is a key research in related research fields in recent years.However,due to the complexity of the causes of brain diseases and the high-dimensional small sample characteristics of current MRI brain image analysis,the performance of classification of brain images is not satisfactory.Under the support of the scientific research project from the Education Department of Hunan Province "research on machine learning methods for spatially structural analysis of brain imaging data",in this paper,we use deep learning methods and brain science background knowledge to study high-dimensional small samples classification problems,and applies them to the classification of Schizophrenia and Alzheimer’s brain images.Our main works are summarized as follows:A comparison experiment of the existing mainstream shallow machine learning methods and basic deep learning models used in the diagnosis of brain diseases on the small sample classification tasks of MRI brain images of Alzheimer’s disease and schizophrenia is conducted,and we obtained benchmark performance of mainstream machine models and basic deep learning models on small sample classification of MRI brain images for these two diseases.We propose a sample generation method based on disease evolution and the pix2 pix generative adversarial network to synthesize new samples and solve the problem of small sample classification of structural MRI(s MRI)brain images in Alzheimer’s disease.Firstly,we augment samples along the linear path of disease evolution by using samples augmentation method which is inspired by prior knowledge of brain science and give them pseudo-labels,then put them into the pix2 pix generative adversarial network to learn the distribution of target dataset from the external MRI dataset to generate new samples,and finally train multiple classifiers on the new samples.We performed small samples classification experiments on Alzheimer’s disease.From the experimental results,we conclude that our new proposed method further improves the accuracy of Alzheimer’s s MRI classification compared to benchmark performance.We propose a deep learning model that combines multi-angle manual feature extraction and deep capsule network to solve the problem of small sample classification of functional MRI(f MRI)brain images in Schizophrenia.The proposed approach proceeds in three steps: extracting linear sparse representation,nonlinear multiple kernel representation,and function connection of brain areas image features from the original 4D brain images;then feeding these image features into three specially designed multilayer deep capsule networks for classification;obtaining the final fusing output results of these three deep capsule network by using a new optimization weighted integration method to realize multi-decision integration.Finally,the proposed approach is implemented and evaluated on the Schizophrenia f MRI dataset.From the experimental results,we conclude that the proposed method outperforms some current benchmark methods in generalization and further improves the accuracy of schizophrenia classification. |
PDF Full Text Request