Deep Feature-Based 3-way Classification of Early to Late Mild Cognitive Impairment in Alzheimer's Disease Using Physics/Chemistry-Based Wrapper Methods for Feature Selection and Machine Learning

Alzheimer’s disease (AD) is the leading cause of neurodegenerative disorders among the elderly. Given its progressive nature, there is a critical need for an early and automated image evaluation system to accurately diagnose this disease. It is crucial to develop extensive, cost-effective, and minimally invasive screening programs, considering the limited availability of scans. Deep learning techniques allow for more efficient development of these systems. This study presents a novel approach for the classification of early (EMCI), mild (MCI), and late mild cognitive impairment (LMCI) using 2D MRI images, utilizing deep features. We obtained three sets of deep features by combining each fully connected layer of the AlexNet, VGG16, and VGG19 models. Next, we investigated five physics/chemistry-based wrapper methods for the feature selection process. The selected features were then classified using six machine learning algorithms: Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), Multilayer Perceptron (MLP), Logistic Regression (LR), and Linear Discriminant Analysis (LDA). Our approach demonstrated high performance, achieving an accuracy of 94.58%, precision of 94.60%, F1 score of 94.58%, recall of 92.47%, Kappa of 88.53%, MCC of 88.60%, and AUC of 98.15%. These performance values were obtained with selected features by Evolutionary Optimization with MLP classifier. The achieved outcome indicates the success of our approach compared to the current state of the art.