A two-level dynamic model for the representation and recognition of cortical folding patterns

We developed a hierarchical framework for the representation of variable compound objects in 2-dimensional images in terms of an adaptive two-level shape model that can be used for recognition and classification tasks. A control instance guides the local search for shapes by managing knowledge about variable spatial relations between single deformable shapes that is uniformly represented through explicit dynamic models. Temporary and permanent memory of past successful searches are used to accomplish a purposeful search, and to improve the a-priori model using generated training data. The framework is applied to the labeling of Heschl´s gyrus in flattened parametric representations of the human cortex (flat maps), which is of great interest with respect to multi-subject fMRI studies. Results indicate the models´ potential for recognition and classification without the need for prior training.