Detecção automática de alterações estruturais hipocampais em imagens de ressonância magnética para auxílio ao diagnóstico da doença de Alzheimer

Abstract

Alzheimer's disease (AD) is the most common form of dementia in world population, accounting for around 60% of all dementia cases and affecting nearly 20% of the population above the age of 80 years. It is an irreversible degenerative disease that causes loss of mental function due to deterioration of brain tissue. This disease is able to develop itself in different ways and its symptoms usually show up gradually. The most common prognostics include the inability to mentally record new information, some degree of difficulty to solve simple problems and to complete familiar tasks at home, confusion about current time and place and trouble to understand visual images. Currently, hippocampus reduction is considered one of the most important, and commonly used AD biomarker. 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 context, this research aims to create an automatic computational technique capable of detecting and classify structural changes in magnetic ressonance (MR) images that differ from AD when compared to cognitively normal (CN) patients. For this, a probabilistic atlas of 3D salient points was built using a dataset of healthy brain images. 3D salient points were detected in the training dataset with CN and mild-AD brain images and used to label each atlas points. Then, the 3D salient points detected in each training dataset image were "matched" against the labeled points in the atlas, and their descriptors vectors were used to train two classifiers, K-NN and SVM-RBF. After that, 3D salient points were detected for each image from the test dataset, and its respective descriptor was used to find equivalent salient points in the atlas. Their descriptors were inserted and classified in K-NN and SVM-RBF classifiers. Finally, each image was labeled accordingly to the majority of points classified in the corresponding class. This technique was tested in different age groups of the ADNI database (with a total of 551 MR images) and the results were evaluated using ROC curve and 10-fold cross-validation. The highest accuracy value achieved by this technique was 85% (up to 82.59% sensitivity and 88.50% specificity) for the 70-89 age group and the highest AUC was 0.9227.