Automatic Labeling of the Human Cerebral Cortex

Glossary 'fsaverage' Subject FreeSurfer subject (Figure 1b and 1c) obtained by averaging 40 subjects. This average subject forms the basis of the FreeSurfer surface coordinate system (Fischl, 2012). Graphical Models The different algorithms in this article can be summarized using graphical models. Graphical models provide a general framework for representing complex probabilistic models with explicit (conditional) independence assumptions. See Figure 2 and 'Summary' section for more details. Markov Random Field Markov Random Fields (MRFs) provides a principal approach to impose a prior on the spatial configuration of labels. One common prior is to favor adjacent spatial locations to have the same labels. See 'Spatially Dependent Priors: Markov Random Fields' section for more details. Mixture Models Registration-based labeling assumes the availability of pre-existing labels. By contrast, mixture modeling only requires that each label generate its own unique image features. Given observed image features, there are algorithms (e.g., Expectation-Maximization) that simultaneously estimate the labels and the probabilistic distribution of image features for each label. See 'Mixture Models' section for more details. MNI152 Template The International Consortium on Brain Mapping (ICBM) average brain obtained by aligning and averaging 152 subjects (Fonov et al., 2011). Registration-based Labeling Registration is the process of establishing spatial correspondences between images. The resulting deformation can be used to transfer pre-existing labels from an (single-subject or probabilistic) atlas to the target brain. See 'Registration-based Labeling' section for more details. The human cerebral cortex is a highly folded 2D sheet of neural tissue that possesses a mosaic of functionally distinct areas oriented parallel to its surface (Figure 1). Information processing proceeds via the transformation of neural signals across these areas (Felleman & Van Essen, 1991; Ungerleider & Desimone, 1986). Accurate labeling of areal locations is therefore an important problem in systems neuroscience. There are four main criteria for distinguishing cortical areas (Felleman & Van Essen, 1991; Kaas, 1987): function, architectonics, connectivity, and topography. Each of these criteria can be interrogated in the human cerebral cortex using a broad range of techniques (Figure 1). Because (1) cortical folds are visible to the naked eye and in anatomical Magnetic Resonance Imaging (MRI) and (2) some cortical folds are indicative of certain underlying cortical areas 2010b), cortical folds are often used as macroanatomical landmarks for comparison of results across brain imaging studies. Consequently, methods that can accurately label cor-tical folds across different subjects are important for studying the human brain (Figure 1). The …