Figure 13 attempts to capture the difference between the two traces over the whole of the head at one level.  The top graph shows, for each of the roughly 1600 network-drawn pixels, the distance to the closest "true" pixel.  The bottom graph is a histogram of the differences.  98% of the pixels are within 2 pixels of the "truth" and 80% are within 1 pixel, an excellent result from our early system testing.  Care must be taken in quantifying this comparison, since some measures of the difference here can be very misleading.  For example, the Hausdorff measure ([4]) of the difference between the two curves is 11, the maximum value of the upper graph; this is almost irrelevant, though, since the measure of this system's success is based on how close we come for the bulk of the pixels, not how far off our single worst difference is.

     Figure 14 is an example of where an operator would override the system, at a bump in the skin (where head restraints were placed).  When the neural network is not well-trained for its task, there will be many operator overrides.  Once captured, though, these exceptions can be added to the training set for the network, and incremental learning will allow the system to improve its performance over time.

     In summary, the general benefits of this system are:

• It can be up to an order of magnitude faster, thus greatly improving a tracer's productivity.
• The system-generated traces are repeatable and consistent across levels; this consistency could be used to address other problems, such as registration.
• It capitalizes on what each do best: the human operator establishes the global context, while the system automates local analysis.
• People can focus on the more interesting, less repetitive parts of task.