Early Detection of Alzheimer’s Disease Using Transfer Learning

Abstract

Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder and the most prevalent cause of dementia worldwide, severely impairing memory, cognitive function, and daily living activities, particularly in the elderly population. Early and accurate detection of AD and its precursor stages is essential
for timely clinical intervention and better patient outcomes. Conventional diagnostic methods relying on manual analysis of Magnetic Resonance Imaging (MRI) scans are time-consuming, subjective, and require large annotated datasets for training deep learning models from scratch. This paper proposes a transfer
learning framework based on MobileNet, pretrained on the ImageNet dataset, to classify brain MRI images into four clinically significant categories: Non Demented, Very Mild Demented, Mild
Demented, and Moderate Demented. The model is trained and evaluated on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset comprising 778 brain MRI images. By reusing pretrained convolutional feature representations and retraining only the final classification layer, the proposed approach achieves high accuracy without requiring large volumes of labeled medical imaging data. With a training configuration of 800 steps, a learning rate of 0.01, and a batch size of 100, the proposed model
achieves a training accuracy of 99.0% and a testing accuracy of 97.56%. These results demonstrate that transfer learning with MobileNet provides a lightweight, computationally efficient, and clinically reliable solution for Alzheimer’s disease detection, suitable for deployment on mobile and embedded platforms.