Neural nets in boundary tracing tasks

Our focus is to use neural networks to interactively assist in the initial segmentation of medical imagery, through learning the characteristics of a contour being traced and projecting ahead a trace whose initial few pixels were specified. To date, much of this work is done manually, since automatic techniques have yielded less than satisfactory results due to prerequisite background knowledge and noise in the data. In our framework, the expert interacts with the network to provide the context, and the network learns the characteristics of the (potentially noisy) locality and continues the task until further guidance is needed. We present here an initial application of this approach to brain MRIs, and we discuss our initial evaluation of neurologically-inspired preprocessing on the input pixel space. Our research directions are discussed