Improve the inputs for more robust performance: We have several ideas for experiments to find a compact and robust set of inputs for the neural network.  Compact input vectors are important to the speed of learning.  Additionally, inputs from adjacent slices may be useful in resolving ambiguous situations on a given slice.

     Extend the algorithm to sub-pixel resolution: Currently, human tracers work against an enlarged image.  This allows the contour to be recorded at fractional pixel resolutions, which thus improves the models generated from the imagery. This initial software, however, works on integer pixel coordinates.

     Study the generality of the learned networks: How general is a network, once learned? The current system stores each learned network with the corresponding set of contours. We would like to study how different these networks really are. There may be a common base of learned cases, to be used either as is, or used to bootstrap the learning of more specific instances.

     Single-platform integration with model building system:  The four steps to generate a surface model, outlined earlier, are currently implemented on different systems.  One goal of future work is to integrate the four pieces into a smoothly interfaced system.