Abstract
Alzheimer's disease (AD) is one of the primary causes of dementia in the older population, affecting memories, cognitive levels, and the ability to accomplish simple activities gradually. Timely intervention and efficient control of the disease prove to be possible through early diagnosis. The conventional machine learning models designed for AD detection work well only up to a certain point. They usually require a lot of labeled data and do not transfer well to new datasets. Additionally, they incur long periods of retraining. Relatively powerful models of deep learning, however, also are very demanding in computational resources and data. In light of these, we put forward a new way of diagnosing AD using magnetic resonance imaging (MRI) scans and transfer learned convolutional neural networks (CNN). Transfer learning makes it easier to reduce the costs involved in training and improves performance because it allows the use of models which have been trained previously and which generalize very well even when there is very little training data available. In this research, we used three different pre-trained CNN based architectures (AlexNet, GoogleNet, and MobileNetV2) each implemented with several solvers (e.g. Adam, Stochastic Gradient Descent or SGD, and Root Mean Square Propagation or RMSprop). Our model achieved impressive classification results of 99.4% on the Kaggle MRI dataset as well as 98.2% on the Open Access Series of Imaging Studies (OASIS) database. Such results serve to demonstrate how transfer learning is an effective solution to the issues related to conventional models that limits the accuracy of diagnosis of AD, thus enabling their earlier and more accurate diagnosis. This would in turn benefit the patients by improving the treatment management and providing insights on the disease progression.