Evaluation of two CNN models, VGGNet-16 & VGGNet-19, for classification of Alzheimer’s disease in brain MRI scans
Abstract: Computer-aided-diagnosis (CAD) emerged in the early 1950s and since then CAD has facilitated the diagnosing of many medical conditions and diseases. In particular, CADfor Alzheimer’s disease (AD) has been immensely researched the last decade thanks to advanced neuroimaging techniques such as magnetic resonance imaging (MRI) and positron emission tomography (PET). Today around 44 million people worldwide have AD and researchers hope to discover accurate ways to detect AD before the symptoms begin. There are currently no validated so-called biological markers (biomarkers) for AD, meaning that there are no reliable indicators that can accurately diagnose AD. However, according to experts, machine learning and neuroimaging is among the most promising areas of research focused on biomarkers and early diagnosis of AD. The state-of-the-art machine learning method for image classification are convolutional neural networks (CNNs). At a recent study at Bharati Vidyapeeth’s College of Engineering and Karunya University, a convolutional neural network VGGNet-16 was used in an experiment in order to correctly classify AD using MRI scans. Experimentation was performed on data collected from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The accuracy using the described method was 95.73% for the validation set. The purpose of this bachelor thesis was to compare two different convolution neural network models: VGGNet-16 and VGGNet-19, comparing their results and performances for classifying AD using MRI scans from ADNI database. Sets of images were elected, some include and some exclude the hippocampus, since AD starts spreading in the hippocampus. Using transfer learning, the CNN models were trained with (a) random validation split, (b) cross validation and (c) different slice range not including the hippocampus. The results of this study show that the models were good at classifying true-negative, which is diagnosing a healthy patient as healthy. Hippocampus seems to be a promising biomarker for AD because experiment (c) achieved a lower accuracy than (a) and (b). In conclusion there is no real statistically proven difference between VGGNet-16 and VGGNet-19. Even then, this thesis showed that simpler CNN architectures can be utilized to classify AD with equally mild success rate on a very limited dataset. The two CNN models’ accuracy were between 66.6- 74.8% for classifying AD depending on the training approach.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)