Detection and Classification of Hippocampal Structural Changes in MR Images as a Biomarker for Alzheimer's Disease

摘要

Alzheimer's disease (AD) is the most common form of dementia, comprising around 60% of all dementia cases and affecting 20% of the population over 80 years of age. AD may affect people in different ways. The most common symptom pattern begins with a gradually worsening ability to remember new information, difficulty to solve problems and perform familiar tasks at home, confusion about time or place, and trouble understanding visual images. Currently, the volume reduction of the two hippocampi is the most used structural magnetic resonance imaging (MRI) biomarker of AD. However, despite its clinical use, hippocampal volume reduction is involved not only in AD but also in other dementias and even in healthy aging. In this study, we propose a new computational framework for the detection and classification of hippocampal structural changes in MR images as a biomarker for AD. First, we built a probabilistic atlas of 3D salient points using a dataset of healthy brain images. Then, we detected 3D salient points in a training dataset with cognitively normal (CN) and mild-AD brain images and used them to label each point on the atlas. Next, the 3D salient points detected in each image from the training dataset were matched against the labeled points in the atlas, and their descriptor vectors were used to train a support vector machine with radial basis function (SVM-RBF). Last, we detected 3D salient points, extracted their descriptor vectors, matched them against the atlas and classified them using the SVM-RBF classifier, for each image from the testing dataset. Finally, we attribute a class label (CN/mild-AD) according to the majority of points classified in the corresponding class. We tested our proposed framework using a stratified age group image dataset (551 MR images in total) and assessed the results using a 10-fold cross-validation and ROC methodology. The highest accuracy value achieved by our method was 85% (up to 82.59% sensitivity and 88.50% specificity) for the age group 70-89, and the highest area under the curve was 0.9227.