PDF version of this document
Models of appearance can be evaluated in a simple principled way. Results presented throughout this paper illustrated the method's applicability to the case of brain model evaluation. Although results have not been shown for other data types, it is known to be a generic method that requires no tweaking and remains robust, regardless of various parameters. By fitting and comparing data sets in a suitable manner, good models can be identified and evaluated quantitatively. For the task, one is better off using robust measures such as the shuffle transform, which is immune to small local discrepancies.
The method was used to evaluate models built by registration and it
is therefore valuable as a benchmark tool. Furthermore, a provision
emerges for validation of registration, assuming that optimal registrations
lead to immaculate models.
Acknowledgements. The project is funded by the EPSRC and
forms a part of the Medical Image and Signal (MIAS) Interdisciplinary
Research Collaboration. The brain data was made available thanks to
the Wellcome Trust and was generously contributed by Paul Bromiley.