| Summary: | Neurodegenerative disease is the term used for a range of incurable and debilitating conditions affecting the human's nervous system. Amongst these conditions, Alzheimer's Disease (AD) is responsible for the greatest burden both for the number of people affected and for the high costs in medical care. The challenges of the disease are related to the subtle symptoms, the increasing pace of disability and the long period of time over which patients will require special care. Recent research efforts have been dedicated to the development of computational tools that can be integrated into the workflow of doctors as a complement to support early diagnosis and targeted treatments. This dissertation aims to study the application of Deep Learning (DL) techniques for the automated classification of AD. The study focuses on the role of PET neuroimaging as a biomarker of neurodegenerative diseases, namely in classifying healthy versus AD patients. PET images of the cerebral metabolism of glucose with fluorine 18 (18F) fluorodeoxyglucose (18F FGD) were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The pre-processed dataset is used to train two Convolutional Neural Networks (CNNs). The first CNN architecture aims to explore transfer learning as a promising solution to the data challenge by using a 2D Inception V3 model, from Google, previously trained on a large dataset. This approach requires a preprocessing step in which the PET volumetric data is converted into a two-dimensional input image which is the input to the pre-trained model. The second approach involves a custom 3D-CNN to take advantage of spatial patterns on the full PET volumes by using 3D filters and 3D pooling layers. The comparative study highlights the performance and robustness of these two models in dealing with the limited availability of the labelled data. The performance of the estimators is evaluated through a cross-validation procedure, giving a score of 83.62% for the 2D-CNN and 86.80% for the 3D-CNN. The results achieved contribute to the understanding of the effectiveness of these methods in the diagnosis of AD. Given the expected margin for improvements, they can be considered promising and in line with the current state of the art.
|
|---|