Cortical Surface Parcellation Using Spectral and Spatial Message Passing on 3D Coordinates

Abstract

The brain consists of many areas that play various roles. Therefore, distinguishing each area of the brain is important in understanding the function or structure of the brain. However, the cortical surface is very complexly wrinkled with sulcus and gyrus, and each subject has a different shape and size. Consequently, existing studies for cortical surface parcellation required a preprocessing process to convert rendered data from cortical surfaces into spherical or spectral domains. This thesis proposes a model that enables segmentation by performing graph convolution directly on original cortical surface without domain transformation. Our model performs end-to-end segmentation tasks using difference vector between nodes to spectral and spatial information of graph defined on the cortical surface. We evaluated our model performance on Mindboggle dataset, 101 manually labeled cortical surface. Compared with previous parcellation methods, our proposed model showed better performance, obtaining the average Dice of 86.25%.