Computer Aided Diagnosis Of Alzheimer's Disease Based On Gaussian Process

Alzheimer's disease is an irreversible neurodegenerative disease.Currently,there are no cures for specific drugs and related technologies.Therefore,early detection of mild cognitive impairment is the only way to save patients.Magnetic resonance imaging is often used in computer-aided diagnosis of such neurological diseases.However,a high data dimension and the lack of training samples have become important factors affecting the recognition rate.The main purpose of this paper is to improve the computer-aided diagnosis performance of AD/MCI based on Gaussian process.The main contents are as follows:1.The paper discusses several commonly used classification methods.And detailed study of the Gaussian process classification model modeling method and solution process.Using the Laplace approximation algorithm under the classification application solves the problem that the posterior distribution is difficult to directly integrate.The classifier performance evaluation indicators used in this paper are elaborated in detail to verify the effectiveness of the improved classification method in this paper.2.On the basis of the Gaussian process,a stepwise classification framework is designed to alleviate this small-sample problem.All test samples are firstly classified by Gaussian process.The samples with high or low posterior probabilities are identified as being correctly classified with high confidence,and then included into the training data to reclassify the rest samples.Experiments on ADNI database show that the second classification tends to increase the posterior probability of the test sample belonging to the right category and improve the classification certainty.The classification performance of the proposed method is superior to the conventional Gaussian process and support vector machine.3.In order to improve the learning ability of Gaussian process classification model for high-dimensional complex data sets,a multi-level Gaussian process classification model was designed based on the idea of the deep belief network model.By inferring any number of Gaussian processes through variational inference techniques.After obtaining the lower bound of variation,the stochastic gradient descent algorithm is used to optimize the parameters.And finally,classify the test samples using the trained multi-level model and Logistic activation function.Using magnetic resonance image data to classify,compare the recognition rates of hidden layers in different quantities to find the optimal hidden layer number.By comparing with the recognition results of conventional Gaussian process classification and support vector machine method,the effectiveness of the multi-level Gaussian process classification method is verified.