Hi Saad,
I guess you mean that the least-squares methods is the same as the maximum
likelihood method if the noise _of the log-transformed data_ is Gaussian,
which we know it is not. Furthermore, even the noise in the "plain" data it
is not Gaussian if the SNR is low, which is many times the case for
diffusion-weighted images using b-values of 1000 or higher.
I don't mean that the least-squares fit to the tensor model is a bad one;
just that you need to know what you are doing.
-Pablo
On Wed, 13 May 2009 17:23:14 +0100, Saad Jbabdi <[log in to unmask]> wrote:
>Hi Megan,
>
>dtifit uses the tensor model. When you log-transform the data, this
>turns up to be a simple linear model which is solved using least
>squares.
>(note that this is the same as maximum likelihood if you assume that
>the noise is gaussian).
>
>If you have FSL4.1.3 or above, you should have access to the --sse
>option, which tells dtifit to output the sum of squares error.
>
>Cheers,
>Saad.
|