Date of Award


Document Type


Degree Name

Master of Science (MS)


Computer Science

Committee Chair/Advisor

Dr. Kai Liu

Committee Member

Dr. Brian Dean

Committee Member

Dr. Nianyi Li


Alzheimer’s disease is an incurable neural disease, usually affecting the elderly. The afflicted suffer from cognitive impairments that get dramatically worse at each stage. Previous research on Alzheimer’s disease analysis in terms of classification leveraged statistical models such as support vector machines. However, statistical models such as support vector machines train the from numerical data instead of medical images. Today, convolutional neural networks (CNN) are widely considered as the one which can achieve the state-of-the- art image classification performance. However, due to their black box nature, there can be reluctance amongst medical professionals for their use. On the other hand, medical images are not easy to get access to, in contrast to general image datasets, such as CIFAR-100, due to several reasons, including privacy and professional cost, motivating us to train the model with high accuracy based on few samples. This thesis focuses on two perspectives: the first interpreting what the CNN model has learned in each layer and will the learned features vary with different input; and second, how to train a reliable network with high accuracy on few medical imaging samples. To address the questions raised above, two different models are examined. First, we use a conventional residual CNN and experiment with two different training methods. The first uses a standard training schedule where the model’s weights are initialized randomly and the second uses transfer learning where we use the weights of a model trained on a larger dataset of a different task as the initial weights for our model. Our method can yield the accuracies of 98.5% and 99.53%, respectively. Our second model studies metric learning instead of classification. In this method the model learns to group images that are similar. The model is fed with a very small set of samples per class, so-called Few-Shot learning. The goal is to learn how similar an image is from another. The model can learn deep embedded representations of an input such that similar inputs are close together in the embedded space and dissimilar inputs are far apart.



To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.