During the last decade extensive research has focused on the potential of neuroimaging for the diagnosis of Alzheimers disease (AD). Currently, the greater challenge is the diagnosis at the early stage known as Mild Cognitive impairment (MCI).

Early detection is important because it is when treatments that can delay the progression of the disease can have the most impact but it is difficult because the spatial pattern of brain degeneration in MCI is highly variable and changes in time as the diseases progresses. Most studies so far used Magnetic Resonance Images (MRI), because they are more widely available and non invasive, and involved a relatively small number of subjects. Although functional imaging modalities such as Positron Emission Tomography (PET) have greater promise in detecting abnormalities useful in early diagnosis and in differentiating patients with MCI that will likely progress to AD, their analysis is even more complex because these images suffer from reduced resolution and poor signal to noise ratio. Therefore, there is a need to develop sophisticated classification tools capable of identifying the spatio temporal patterns which characterize AD and MCI in functional images as well as providing quantitative information of relevance for the clinicians.

The goal of this project is twofold:
1. to get a deeper insight into the manifestations of Mild Cognitive Impairment (MCI) and Alzheimer ‘s disease (AD) both in space and time dimensions
2. to produce more accurate diagnosis tools using PET images of the brain.

In order to achieve these goals, we propose to develop automatic classification tools for PET brain images to:
1. differentiate between Alzheimer’s disease (AD), Mild Cognitive Impairment (MCI), and normal control subjects (NC).
2. distinguish between MCI patients who convert to AD (MCI-C) from those who remain stable, the non-converters (MCI NC).
3. identify individual patterns of disease evolution, by assessing morphological changes along the time, based on follow up scans for each patient.
4. incorporate complementary sources of information, namely based on cognitive evaluation tests, into the diagnosis process.

This research is possible at the present given that a large amount of clinical and imaging data (labeled as AD, MCI and NC) has recently become available from the Alzheimers Disease Neuroimaging Initiative (ADNI) database (www.loni.ucla.eduADNI).

Since a brain volume contains thousands of voxels which represent variables and the number of subjects is generally smaller, this task suffers from the so called ‘curse of dimensionality’. In these cases, traditional classifiers like Nearest Mean Classifier or neural networks are not appropriate. We will investigate state of the art techniques, based on mixtures of classifiers (ensemble methods), such as Boosting and Random Forests. Moreover, for comparison purposes we will also apply kernel methods used recently such as Support Vector Machines (SVM). The above mentioned ensemble methods, besides being more robust to the ‘curse of dimensionality’, determine how important variables are to the classification task, therefore performing an automatic feature selection. This information
is extremely valuable not only from the point of view of clinical interpretation but also to evaluate the possibility of developing classifiers which do not depend either on manual or semi-automatic definition of regions of interest (ROI) or even on prior data reduction techniques such as PCA or ICA.

Besides studying the spatial patterns of individual PET brain images, we also propose to analyze the patterns time evolution both on individual and group basis. Therefore, in addition to single PET scans we will use follow up scans to determine if the simultaneous use of spatial and temporal features can improve the classifiers ability to diagnose individual patients.

Additionally, we propose to incorporate complementary sources of information, namely based on cognitive evaluation tests, into the diagnosis process. This information will be integrated into the classification process exploring supervised, unsupervised and semisupervised techniques.

Our team involves image processing and pattern recognition experts both in the field of supervised and unsupervised methods as well as a neurologist specialist in the field of AD and PET. We expect that the proposed research will contribute with knowledge and methods toward an early diagnosis of Alzheimers disease.