Alzheimer's disease classification based on graph kernel support vector machines constructed with 3D texture features extracted from magnetic resonance images

Alzheimer's disease (AD) is a neurodegenerative disease characterized by cognitive and behavioral impairment that significantly interferes with social and occupational functioning. Mild cognitive impairment (MCI) is a relatively broad clinical condition involving a slight memory deficit, which in many cases represents a transitional state between a cognitively normal (CN) condition and AD. Structural magnetic resonance (sMR) imaging has been widely used in studies related to AD because it provides images with excellent anatomical details and information about structural and contrast changes induced by the disease in the brain. Many published studies restrict their analysis to a few particular regions of the brain and search for structural changes caused by the disease. Recent studies start looking for new AD biomarkers using multiple brain regions and focusing on subtle texture changes in the image. Therefore, this study proposes a new technique for MR image classification in AD diagnosis using graph kernels constructed from texture features extracted from sMR images. In our method, we first segment the MR brain images into multiple regions with the FreeSurfer. Then, we extract 22 texture features using three methods and define the graph-node attributes as the probability distributions of the extracted features. Next, for each texture feature, we build a graph and define its edge weights as the distances between pairs of node attributes using three distance metrics. After that, we use a threshold-based approach for graph edges removal and create the graph-kernels matrices. Finally, we perform image classification using Support Vector Machines (SVMs) with two graph-kernels. Results of our method have shown better performances for the CNxAD (AUC = 0.92) and CNxMCI (AUC = 0.81) classifications, and worse for the MCIxAD case (AUC = 0.78). This trend is consistent with other published results and makes sense if we consider the concept of Alzheimer's disease continuum from pathophysiological, biomarker and clinical perspectives. Besides allowing the use of different texture attributes for the diagnosis of Alzheimer's, our method uses the graph-kernel approach to represent texture features from different regions of the brain image, which considerably facilitates the image classification task via SVMs. Our results were promising when compared to the state-of-the-art in graph-based AD classification.