A Comparative Analysis of Deep Learning Approaches for Alzheimer's Disease Diagnosis Using MRI Images

Alzheimer's disease (AD) is a progressive neurological disorder that severely impacts cognitive function. Early detection and accurate diagnosis are crucial for improving patient outcomes. This study explores the application of deep learning models, specifically Inception V3 and AlexNet, for classifying MRI images into Alzheimer's severity classes: Mild Demented, Moderate Demented, NonDemented, and Very Mild Demented. Inception V3 and Alex net are lightweight models compared to other models (ResNet, VGG16, Efficient Net). These two models assess trade-offs between performance and Computational cost. A comprehensive dataset visualization revealed significant class imbalances, with distributions ranging from 1 % to 50 %. Data augmentation techniques such as shifts, rotations, and zooming were employed to address this, achieving a balanced class representation of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$24-26 \%$</tex>. The models were evaluated on both original and augmented datasets to assess the impact of augmentation on performance. Inception V3 achieved 98% accuracy, 0.99 precision, 0.98 recall, and 0.98 F1-score, outperforming AlexNet, which reached 93% accuracy, 0.92 precision, 0.90 recall, and 0.90 F 1 -score. The study highlights that augmentation enhances sensitivity. Unlike prior studies that predominantly use a single deep learning approach, this research employs a comparative framework, evaluating both lightweight (AlexNet) and deep (Inception V3) architectures. Additionally, the effect of augmentation on class balance and model performance has been thoroughly analyzed, highlighting potential trade-offs in sensitivity and specificity. Despite improved sensitivity and reduced misclassification errors in augmented datasets, introducing synthetic patterns affected generalization. The study highlights the critical role of data augmentation in addressing class imbalances and enhancing model training while emphasizing the need for augmentation strategies that closely mimic real-world data variability.