![]() | Only 14 pages are availabe for public view |
Abstract Alzheimer’s disease (AD) is the most prevalent type of dementia of the nervous system causing weakness of many brain functions, such as memory loss. Non- invasive early diagnosis of AD has attracted a lot of research attention nowadays as early diagnosis is essential for improving the patient care and treatment plan. This thesis develops a deep learning-based pipeline for accurate diagnosis and stratification of AD stages. The proposed analysis pipeline utilizes shallow Convolutional Neural Network (CNN) architecture and 2D T1-weighted Magnetic Resonance (MR), and 18F-fluorodeoxyglucose positron emission tomography (18F-FDG PET) brain images. Classical Machine Learning (ML) algorithms were applied to clinical dataset like (Logistic Regression (w/ imputation), Logistic Regression (w/ drop), Support Vector Machine (SVM), Decision Tree, Random Forest, and AdaBoost). These methods were used to analyze the measures of the patients and make a relation between its risk factor, such as age and personal history to classify into AD patient or normal. The T1-weighted cross-sectional Magnetic resonance imaging (MRI), longitudinal MRI data from a clinical data collection, and 18F-FDG PET brain scans are different publicly available datasets were used to evaluate the performance of the proposed model. The proposed pipeline not only introduces a fast and accurate AD diagnosis module but also provides a global classification (i.e., normal vs. Mild Cognitive Impairment (MCI) vs. AD) as well as local classification. The latter deals with an even more challenging task to stratify MCI into a Very Mild Dementia (VMD), mild dementia (MD), Moderate Dementia (MoD), Early Mild Cognitive Impairment (EMCI), and Late Mild Cognitive Impairment (LMCI) as the prodromal AD stage. In addition, the experimental results were compared to cutting-edge deep learning architectures, such as DenseNet121, ResNet50, VGG 16, EfficientNetB7, and InceptionV3. The reported results established superior accuracy compared to the well-known classification techniques with, overall testing accuracy of 99.68%. |