Abstract
In machine learning, one of the most popular deep learning methods is the convolutional neural network (CNN), which utilizes shared local filters and hierarchical information processing analogous to the brain's visual system. Despite its popularity in recognizing two-dimensional (2D) images, the conventional CNN is not directly applicable to semi-regular geometric mesh surfaces, on which the cerebral cortex is often represented. In order to apply the CNN to surface-based brain research, we propose a geometric CNN (gCNN) that deals with data representation on a mesh surface and renders pattern recognition in a multi-shell mesh structure. To make it compatible with the conventional CNN toolbox, the gCNN includes data sampling over the surface, and a data reshaping method for the convolution and pooling layers. We evaluated the performance of the gCNN in sex classification using cortical thickness maps of both hemispheres from the Human Connectome Project (HCP). The classification accuracy of the gCNN was significantly higher than those of a support vector machine (SVM) and a 2D CNN for thickness maps generated by a map projection. The gCNN also demonstrated position invariance of local features, which rendered reuse of its pre-trained model for applications other than that for which the model was trained without significant distortion in the final outcome. The superior performance of the gCNN is attributable to CNN properties stemming from its brain-like architecture, and its surface-based representation of cortical information. The gCNN provides much-needed access to surface-based machine learning, which can be used in both scientific investigations and clinical applications.