Deep Learning on Brain Cortical Thickness Data for Disease Classification

Deep learning has been applied to learn and classify brain disease using volumetric MRI scans with an accuracy approaching or even exceeding that of a human expert. This is typically done by applying convolutional neural networks to slices of a 3D brain image volume. Each slice of the brain volume, however, represents only a small cross-sectional area of the cortical layer. On the other hand, convolutional neural networks are less well developed for 3D volumes. Therefore we sought to apply deep networks to the 2D cortical surface, for the purpose of classifying Alzheimer's disease (AD). AD is known to affect the thickness and geometry of the cortical surface of the brain. Although the cortical surface has a complex geometry, here we present a novel data processing method to feed the information of an entire cortical surface into existing deep networks for more accurate early disease detection. A brain 3D MRI volume is registered and its cortical surface is flattened to a 2D plane. The flattened distributions of the thickness, curvature and surface area are combined into an RBG image which can be readily fed to existing deep networks. In this paper, the ADNI dataset of brain MRI scans are used and flattened cortical images are applied to different deep networks including ResNet and Inception. Two pre-clinical stages of AD are considered; stable mild cognitive impairment (MCIs) and converting mild cognitive impairment (MCIc). Experiments show that using flattened cortical images consistently leads to higher accuracy compared to using brain slices with the same network architecture. Specifically, the highest accuracy of 81% is achieved by Inception with flattened cortical images, as compared to 68% by the same network on brain slices and 75.9% accuracy by the best method in the literature which also used a deep network on brain slices. Our results indicate that flattened cortical images can be used to learn and classify AD with high accuracy.