Transfer Learning for Alzheimer's Disease Detection on MRI Images

In this paper, we focus on Alzheimer's disease detection on Magnetic Resonance Imaging (MRI) scans using deep learning techniques. The lack of sufficient data for training a deep model is a major challenge along this line of research. From our literature review, we realised that one of the current trends is using transfer learning for 2D convolutional neural networks to classify subjects with Alzheimer's disease. In this way, each 3D MRI volume is divided into 2D image slices and a pre-trained 2D convolutional neural network can be re-trained to classify image slices independently. One issue here, however, is that the 2D convolutional neural network would not be able to consider the relationship between 2D image slices in an MRI volume and make decisions on them independently. To address this issue, we propose to use a recurrent neural network after a convolutional neural network to understand the relationship between sequences of images for each subject and make a decision based on all input slices instead of each of the slices. Our results show that training the recurrent neural network on features extracted from a convolutional neural network can improve the accuracy of the whole system.