Krashenyi I. A fuzzy logic based tomography image analysis method for automated Alzheimer's disease diagnosis

The new method of analysis of tomographic brain image by means of fuzzy logic in this work was developed and investigated. The method of selection and sorting characteristics based on statistical criteria and principal components analysis was developed. The internal structure and principles of operation of the automated system for diagnosis of Alzheimer's disease (AD) based on tomographic brain image analysis by means of fuzzy logic was presented. The experimental studies result of real tomographic images of the human brain in three groups: healthy patients, with AD and with mild cognitive impairment (early stage Alzheimer's) confirmed the efficiency of created and improved methods and means of diagnosis of AD. The proposed approach of analysis based on fuzzy logic using MRI, PET brain images features as well as multimodal features based on MRI and PET brain image for AD computer-aided diagnosis was presented. The proposed approach uses clustering algorithm for rule synthesis. Mean values of voxel intensity in spatial regions of interest which were extracted from normalized MRI or/and PET scans of brain gray matter were used as features. In order to improve the classification performance and to diagnose AD, outputs bagging (averaging) was performed. Tomographic images of the brain that are used in this study were obtained from a database of Alzheimer's Disease Neuroimaging Initiative (ADNI). This image contains 1.5T / 3.0T MRI and PET 18F-FDG images, which were recorded from 249 subjects, of which 68 are control group (NC), 70 represent a group of patients (AD) and 111 represent a group of people with cognitive disorders (MCI). Area under receiver operating characteristic was used as a classification performance measure, being function of the number of brain anatomical and functional regions of interest from which the features were extracted. Leave-one-out cross-validation was used to estimate performance of computer-aided system for AD and mild-cognitive impairments, resulting in accuracy, sensitivity, specificity and positive predictive value characteristics of fuzzy classification between different groups. To estimate the threshold of diagnosis, harmonic mean of sensitivity and specificity as well as area under the ROC-curve were developed in this work. To reduce possibility of over-diagnosis and to improve statistical indicators of the performance of automated diagnosis of AD in this work the methods of obtaining features of tomographic images of the human brain by determining the most discriminant regions of interest in tomographic images of the brain and calculating the characteristics based on anatomical atlases and statistical criteria were further developed. The priority area determination is based on the anatomical atlas by the spatial arrangement of regions of the brain, and based on Kolmogorov-Smirnov, Student's and Mann-Whitney statistical criteria. This approach allows to sort the region of interest in accordance with the difference between them for tomographic images of healthy and sick brain and determine the most discriminant region to account only the most significant changes in tomographic images of the brain. On the basis of the principal component analysis and tomographic images voxel reorganization the feature extraction method was improved. For this purpose to calculate an ordered basis of eigenvectors by solving equations and eigenvectors sorting according to values of corresponding eigenvalues modules.