Advancing Dementia Classification Through Intelligent Machine Learning and Deep Learning Approaches

Abstract

Despite significant advances in understanding the structure of dementia through structural magnetic resonance imaging (MRI), classification of dementia from MRI data is still challenging due to the high dimensionality of features, the redundancy of information, the varied patterns of dementia, and the limited generalizability of models. This study presents an intelligent framework that combines the metaheuristic optimization, machine learning (ML), and deep learning (DL) techniques to achieve automatic dementia classification. A novel Adaptive Lévy–Opposition Search Optimization (ALOSO) algorithm is presented to improve on global exploration, local exploitation, feature selection, and model optimization. The framework assesses the performance of Support Vector Machine (SVM), Random Forest, XGBoost, Convolutional Neural Network (CNN) and three-dimensional CNN (3D-CNN) models using MRI-derived morphometric and deep image representations. The model performance on classification is evaluated by Accuracy, Precision, Recall, F1-score, Specificity and ROC-AUC, and ablation and independent cross-dataset validation are added to explore the robustness of the model. ALOSO reduced the feature space from 120 to 43 variables, and the proposed hybrid model achieved the accuracy of 96.0%, the F1-score of 95.3% and the ROC-AUC of 98.0% for illustrative experimental setup. The framework offers a proof of concept for the use of optimized anatomical features and learned spatial representations for robust dementia classification.