Discussion

The derived similarity function, $F_0$, demonstrates superior accuracy and robustness when compared with the simpler similarity functions, $F_1$ and $F_2$. The partial volume terms decrease the number and extent of discontinuities, and the terms that normalise for the number of model parameters, $N_{eff}$, de-weight erroneous local maxima. In addition, this similarity function automatically includes prior shape information, via $p(T)$, as well as partial volume and bias field effects without needing adjustable, ad-hoc parameters.

Future work will apply this similarity function to our intended application of anatomical shape segmentation of the human brain and extend the above to incorporate intensity priors, and modelling intrinsic tissue parameter distributions with one variance per shape.