After the evaluations done by Klein et al (published in NeuroImage), I
figured that the use of slightly more regularisation may increase the
accuracy slightly. Klein's evaluations seemed to show that the
algorithm was aligning GM with GM with a very good overlap of tissue
type, but that its extrapolation of the deformations into other regions
may not be as accurate as it could have been. A slight decrease in the
amount of regularisation should therefore improve the extrapolation, and
the overall accuracy.
In practice, the amount, and form, of regularisation used by the
registration model will often have a bigger effect on the resulting
accuracies than the differences between the actual implementations
themselves. Also, because of inherent bias-variance trade-offs, the
optimal amount of regularisation will depend on the aims of the study,
and possibly how many subjects are included.
On Wed, 2009-11-11 at 20:10 -0800, Fumiko Hoeft wrote:
> Dear experts,
> I was wondering if one discrepancy we found between SPM5 and SPM8
> dartel is a bug or feature. The section we are referring to is the
> regularisation parameters of the last outer iteration. SPM5 is "0.125
> 0.0625 1e-06" but in SPM8, the default is "0.25 0.125 1e-06". Could
> someone let me know?
> Thank you.
> Fumiko Hoeft MD PhD
> CIBSR, Stanford Univ Sch of Med
> a- 401 Quarry Rd. M/C 5795, Stanford CA 94305-5795
> t- (650)245-7016 f- (650)724-4794
> e- [log in to unmask]
> u- http://cibsr.stanford.edu http://stanford.edu/~fumiko http://readingbrain.stanford.edu
> CONFIDENTIALITY NOTICE: Information contained in this message and any
> attachments is intended only for the addressee(s). If you believe that
> you have received this message in error, please notify the sender
> immediately by return electronic mail, and please delete it without
> further review, disclosure, or copying.
John Ashburner <[log in to unmask]>