Dear Juan,
For display purposes, you could use the unthresholded SPM{t} map
(spmT_xxxx.nii) or contrast image (con_xxxx.nii). Now, if the quantity
of interest you are estimating with your model is comparable to a
contrast value from SPM's GLM then you could enter all of the data from
the two models and the two groups in a factorial design and look for a
group-by-model interaction in order to highlight regions benefiting from
your approach.
Best regards,
Guillaume.
On 22/01/2020 11:43, Juan A. Arias wrote:
> Dear SPM users,
>
> I'm doing a Ph.D. in Neuroscience with a special focus on improving statistical methods for early diagnosis in Alzheimer's disease. My first goal is to implement R code for visualization of PET images with advanced non-parametric regression models (i.e. Generalized Additive Models). The outcome of this code is an image which predicts which areas have the same activation patterns and then propose iso-activity lines (see attached image "vis.gam contour", work in progress).
>
> These images are very precise, however, my interest is to compare the utility of these outcomes with the ones extracted from SPM so that I can evaluate the potential progress that my package would provide to the field. And that's where the problems begin. First, SPM outputs consist on images with highlighted voxels/regions where PET activity is significantly higher (or lower in this case as the contrast is between Controls and Alzheimer patients so I'm looking for hypometabolism) but I can't seem to figure out how to extract a "model" of activity patterns to compare outcomes. How can I compare Image 1 with Image 2 attached (for example)?
>
> Three approaches come to my mind:
>
> -1) Somehow (I guess using the residuals written by SPM after the statistical contrast) I should be able to calculate the amount of "lost information" by SPM and compare it with the amount of information lost with my regression models in R. There are several ways of comparing this loose of information (AIC methods for example) with regression models but I can't figure out how to get my two sources of data to be comparable between them. However, I guess there has to be a way to do that.
>
> -2) Somehow create models of brain activation with SPM (something similar as Image 1), although I can't find any tool or information online on how to do this, and then compare these models with the ones retreived from R.
>
> -3) Going a step forward with R, compare regression models obtained from Control Group vs regression models from Alzheimer's Group, carry out T-tests to extract the areas with significant hypometabolism, and then my outcome in R would be something more similar to a Parametric Map. Still, I'm unsure on how to compare efficiency between the two approaches. How can I have a general overview of the accuracy of both processes?
>
>
> Any ideas on how to proceed, journal articles, books...or any other source of information on this problem would be gratefully acknowledged!
>
>
> Thank you so much in advance.
>
>
> Juan A. Arias
> Clinical Neuroscience Ph.D. Student
> Biostatistics and Biomedical Data Science Unit
> university of Santiago de Compostela (Spain)
>
--
Guillaume Flandin, PhD
Wellcome Centre for Human Neuroimaging
UCL Queen Square Institute of Neurology
London WC1N 3BG
|