Abstract
INTRODUCTION: Alzheimer's disease (AD), the most prevalent form of dementia, affects more than 50 million individuals worldwide and demands accurate and timely diagnosis to improve patient outcomes. Traditional machine-learning approaches for AD detection using MRI often rely on manual feature extraction, which is labor-intensive and limits scalability. There is a growing need for automated, high-accuracy methods that can support clinical workflows and respond to the expected tripling of AD cases by 2050. METHODS: This study proposes an automated feature-extraction approach using a pre-trained ResNet50 convolutional neural network (CNN) applied to brain MRI scans. Extracted deep features were classified using three different algorithms: Softmax, Support Vector Machine (SVM), and Random Forest (RF). Performance was evaluated on two benchmark datasets: ADNI and MIRIAD. RESULTS: Among the tested models, the ResNet50-Softmax combination demonstrated the highest performance, surpassing state-of-the-art benchmarks (85.7%-98.59%). It achieved 99% sensitivity, 98% specificity, and an overall 99% accuracy on the ADNI dataset. On the MIRIAD dataset, the model also performed strongly, reaching 96% accuracy. DISCUSSION: The results confirm that transfer learning using ResNet50 significantly enhances the accuracy and scalability of AD diagnosis from MRI data. By eliminating the need for manual feature extraction and offering near-perfect classification performance, this approach can streamline clinical neuroimaging workflows. These findings highlight the potential of deep learning models to support early diagnosis and meet the increasing global burden of Alzheimer's disease.