Hello all,
I was wondering if anyone could help me understand how to appropriately extract beta coefficients from a first-level FSL feat to perform a classification analysis.
I have a working pipeline that uses SPM to extract trial-wise betas for classification. However, my data has been analysed with FSL, and includes ICA and voxel-wise PNM regressors. I can add the ICA regressors to the SPM GLM, but I do not know how to incorporate the voxel-wise regressors from PNM. Therefore, I have been attempting to classify using the pe.nii.gz or beta coefficients from the FSL analysis instead (this would also prevent me from having to analyse all of my subjects again in SPM).
The problem is, when I classify using the FSL betas, I am getting below chance accuracies (33%), whereas with the SPM pipeline, I am getting mean accuracies of 98%. Therefore, I must be extracting the FSL betas incorrectly.
I am using the pe.nii.gz files from the first level feat, but I have transformed them from functional to structural space and lowered the resolution. I read online that FSL applies grand mean scaling that prevents comparison of betas across runs? Therefore, I tried to scale the betas using the following command:
fslmaths pe_img -mul scale_factor -div mean_func -mul mask_image output_image
But I still have below-chance classification accuracy using this method.
Any advice or tips regarding extracting betas from FSL for classification or including voxel-wise PNM regressors into the SPM GLM would be amazing!
Thank you for reading,
Paige