University of Ioannina
Αυτόματη ανάλυση λειτουργίας εγκεφάλου σε άτομα με τη νόσο Alzheimer βασισμένη στον αλγόριθμο τυχαίων δασών
Abstract
dc:descriptionThe aim of this thesis is the development of an automated method for the analysis of the brain function in patients with Alzheimer’s disease (AD). Based on the processing and analysis of functional Magnetic Resonance Imaging (fMRI) and on the utilization and improvement of machine learning techniques, we propose a method for the diagnosis of AD and the separation of stages of the disease progression. The proposed method consists of six steps. First, the functional Magnetic Resonance Images were preprocessed, in spatial and time domain, to remove artefacts and the subsequent analysis to lead to reliable results (Step 1). The voxel time courses were extracted from the preprocessed images and they were modeled using a generalized linear model (Step 2). The result of the second step is the formation of a statistical parametric map for each subject. The activation maps were analyzed permitting thus the extraction of features describing the activation patterns and the hemodynamic response of the brain. The functional images were also used for the calculation of an index expressing the head movement during the conduction of the fMRI experiment. Additionally, the anatomic MRI images were exploited to extract features describing the atrophy of gray matter (Step 3). Those features, as well as, demographic details (age and sex) and behavioral data (median and average reaction time) were evaluated for their contribution to the diagnosis of AD and the monitoring of its progression. For this purpose a feature selection approach, based on information theory measures, was applied (Step 4). The selected features were supplied as input to the classification step (Step 5). The Random Forests algorithm and Support Vector Machines (SVMs) were evaluated for their classification accuracy, sensitivity and specificity. Finally, the trees that consist of the forest were converted to rules (Step 6). The rules combine the information, which is encompassed to the features, and they provide the decision if the subject is suffering from AD. In case of positive diagnosis the rules determine the stage of progression (very mild or mild). The method was evaluated using data from 41 subjects (14 healthy young, 14 healthy old, 13 demented old subjects). The data were made available at the fMRI Data Center (fMRIDC) by Buckner and his colleagues. The classification results, for the diagnosis of the disease and the progression of its stages, indicated that the Random Forests algorithm outperforms the Support Vector Machines. This can be attributed to the fact that Support Vector Machines are a single classification model while Random Forests is an ensemble classifier. Several authors, in the literature, have demonstrated the advantage of the ensemble methods over the individual classifier models and have noted that they significantly improve the performance of the learning. Taking into consideration the above issues we propose modifications of the Random Forests algorithm which improve its prediction performance. The suggested modifications intend to increase the strength and decrease the correlation of individual trees (two factors that affect the generalization error of the algorithm) and to improve the function which determines how the output of the base classifiers is combined. This is achieved by modifying the node splitting and the voting mechanisms. The novel feature of this thesis is that proposes modifications, not only for the improvement of the construction or the voting procedure but, for the first time, it examines the overall improvement of the Random Forests algorithm (a combination of construction and voting). For the evaluation of the proposed modifications 14 datasets were used. Thirteen of them are publicly available at the UCI Machine Learning Repository. The datasets cover different areas of science and they consist of a different number of features and classes. We compare the proposed modifications in terms of accuracy and other heuristics reported in the literature. Also, a statistical comparison was conducted to examine if there is a statistical significant difference between the results of the above modifications. The study of the Random Forests algorithm revealed that a significant parameter of the algorithm and for the ensemble classifier in general, is the number of base classifiers which consist of the ensemble. In this thesis we develop a method for the dynamic determination (during the construction of the forest) of the number of base classifiers. The method is based on an on-line fitting procedure. The fMRI technique has been employed in many studies in the literature concerning patients with neurodegenerative diseases such as AD. Most of the studies are indicative of how the specific technique can be used as a diagnostic tool for different neurological abnormalities. However, the studies reported in the literature focus on detecting differences between different healthy and demented subjects using fMRI. Only few of them present an index for the diagnosis of AD. They are concentrated on a specific anatomical region depending thus on the presentation of a specific task for the activation of this region. In addition, they presuppose a specific type of fMRI experiment (block design, event related, resting-state) and they are using only one feature, activation, in order to extract conclusions. On the contrary, the method proposed in this thesis not only diagnoses the AD but it specifies the stage of progression. It provides an explanation of the results through the rules that are extracted and it does not limited to detect differences between group of subjects but quantifies these differences through a number of features expressing a big part of the medical knowledge about AD. This is achieved by exploiting all the information provided by the fMRI experiment. Finally, it is independent form all the factors mentioned above and it can be easily adapted to the diagnosis of a new type or the differentiation between different types of dementia.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2012
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tripoliti, Evanthia
- Τριπολίτη, Ευανθία
Subjects
dc:subject × 12Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/26803