Alzhei̇mer's disease detection using ensemble learning on brain MRI images

Alzheimer's disease is a progressive neurodegenerative disorder that affects millions of people worldwide, gradually impairing an individual's cognitive, behavioral, and motor abilities. As the most common form of dementia, it develops due to the damage or death of neurons and synapses in the central nervous system, leading to a decline in essential cognitive functions such as memory, language, and decision-making. In the early stages, symptoms such as forgetfulness, inattention, and confusion are often mistaken for age-related or psychological conditions, making timely and accurate diagnosis difficult. The disease is characterized by the abnormal accumulation of beta amyloid plaques and tau proteins in the brain, which damage nerve cells and lead to structural shrinkage and tissue degeneration. Magnetic Resonance Imaging (MRI) plays a critical role in detecting these structural changes by monitoring reductions in brain volume, tissue atrophy, and degeneration in specific regions. However, in the early stages of Alzheimer's, such changes are often subtle and ambiguous, making classical image-based assessment methods insufficient. Moreover, traditional diagnostic procedures are time-consuming, costly, and subjective, as they depend heavily on expert interpretation. In recent years, Artificial Intelligence (AI), particularly Deep Learning-based approaches, has demonstrated remarkable success in medical imaging analysis. Convolutional Neural Networks (CNNs), known for their high capability in learning spatial patterns from images, have been effectively utilized in classifying Alzheimer's stages. To overcome the limitations of individual models and enhance overall performance, Ensemble Learning techniques that combine the strengths of multiple CNN architectures have gained increasing attention. In this study, the stacking ensemble method was employed to integrate outputs from several CNN models through a meta-learner to improve generalization and diagnostic accuracy. The experiments utilized an open-access MRI dataset from Kaggle, representing four stages of Alzheimer's disease (Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented). The dataset was balanced across training, validation, and test sets. Pre-trained CNN architectures AlexNet, DenseNet, ResNet, GoogleNet, EfficientNet, MobileNet, and VGG16 were fine-tuned through transfer learning. Each model's performance was evaluated using accuracy, F1-score, AUC-ROC, and log loss metrics. Their outputs were subsequently combined into a meta model using stacking to enhance classification performance. To further increase the interpretability and transparency of the decision-making process, the Grad-CAM (Gradient-weighted Class Activation Mapping) technique was applied. Grad-CAM visualizations revealed that the models primarily focused on Alzheimer-related brain regions such as the hippocampus and temporal lobes, xx demonstrating that the system produces decisions that are not only statistically sound but also neurologically meaningful. The results indicated that CNN models employing transfer learning outperformed models trained from scratch, while the ensemble approach achieved higher accuracy compared to individual CNN architectures. The proposed system achieved approximately 98% classification accuracy, confirming its potential as a reliable and explainable decision support tool for medical image analysis. In conclusion, the study offers an innovative contribution to the early diagnosis of Alzheimer's disease, combining technical robustness with clinical applicability.