PDF version of this document
The simple data used by Smith is already proving slightly too cumbersome for responsive experimentation on a relatively strong machine (1.8 GHz, 512MB RAM), especially owing to the complex algorithms devised for group-wise registration. It is advisable that evaluation via profiling toolkits is firstly made to hasten this process as much as possible. Alternatively, coding of the algorithm in a compiled language as C++ is seriously looked at as a possibility. The complexity of the departmental VXL library is believed to make such step less than desirable.
Once speed-up has been taken care of or when it is at least known that a near-flawless well-performing piece of software is at disposal and under control, the simple 1-D data can see the addition of a few additional characteristics. This new composite46 data must retain some good commonality and similarity across the set of images and it should not be overly more complex and unpredictable in comparison with a simple bump. A double-humped curve, a round smooth line or even a contour of a a profile of a face might be sensible and more challenging choices47. In any case, whichever synthesis of data is eventually selected and experimented with, the choice of control points for the warping then becomes a more crucial issue48. A more localised control via warps then turns mandatory because several separate structures exist in the data.
The experiments of Marsland, Twining and Taylor have already shown the realistic application of warps to a medium-resolution two-dimensional data. Nonetheless, it is vital to point out that an elliptical shape was dealt with and a priori knowledge of the problem was used to increase the speed of the group-wise registration process. Control points that characterise the warps were initially places on a circle whose centre was the image centre and radius corresponded to the typical position of the skull in standardised imaging. If the problem involved point selection for, let us say, knee cartilage and no knowledge about the object was available in advance, the results would have then taken far more than 10 hours to obtain (as was the case for the 12 points distribution around the skull's exterior). Edge detection is quite useful in an application of this kind. It was highly useful in the case of the skull data, but finding edges that form a circle (confer Hough transform) as in a skull is somewhat of a simplified problem. Developments should aim to address many such issues.