COMPUTERIZED DIAGNOSIS OF ALZHEIMER’S DESEASE FROM FUNCTIONAL MAGNETIC RESONANCE

Background and objective: AD is the most common kind of dementia affecting millions of people worldwide. Its severity and social impact is given by the consequences related to this pathology, affecting patient and family lives. Several studies in the last decades have focused their attention of the understanding of this disease and proposing faster and more accurate diagnosis. The purpose of this thesis is to explore this pathology and the state of the art of the Machine Learning (ML) algorithms implied in this field. Moreover, starting from the research done in a recently-published study where they propose a 3D framework called Brain-on-Cloud for the identification of AD from structural MRI (sMRI) scans, the main purpose of this thesis is to verify the effectiveness of Brain-on-Cloud in pursuing the same classification task by taking into account a different data type from sMRI, precisely functional magnetic resonance images (fMRI). Methods: In order to have a solid background for the understanding of the disease, the anatomy and physiology of the brain are analysed in detail, being the most complex organ in the human body. AD consequences in the brain are explored and include tissue damage and impairment of neuronal communication, leading to cognitive and physical issues, which worsen in time as a degenerative pathology. The diagnostic techniques in use for AD are described, focusing the attention on Magnetic Resonance Imaging (MRI) principles and application. After the study of the basic principles of ML and its use in medicine, the state of the art of its application in AD diagnosis is explored. To the aim of verifying the effectiveness of Brain-on-Cloud on fMRI images, the first experiment of the reference study is implemented, including the basic steps of the image pre-processing: intensity normalization and image cropping. Moreover, the model hyper-parameters are investigated, the best hyper-parameter combination is selected and used, and the results of the first experiment of Brain-on-Cloud by using both sMRI and fMRI scans are compared and discussed. Results: The best combination of hyper-parameters obtained is the following: batch size of 10, learning rate equal to 0.005 and dropout rates of 0.5 and 0.6. With the implementation of this set of hyper-parameters, a mean area under the curve of about 0.86 is obtained and is comparable with the one in the reference study. Moreover, this model is able to identify AD with an accuracy of 77.45%, sensitivity of 81.08 % specificity of 73.73 % and F1-score of 77.95 %. Conclusion: In this thesis only a preliminary evaluation is made on fMRI scans, implementing the first experiment performed in the reference study. The promising results obtained open the perspective of the use of this kind of images as input data of Brain-on-Cloud for AD identification, obtaining comparable or even better performances with respect to the already tested structural magnetic resonance images (sMRI). The outcome of this research also highlights the importance of hyper-parameter tuning as necessary tool for the improvement of the model.