MRIQC: any experience with custom classifier for func images?

1

Hi, I believe that MRIQC does not have a classifier for func images yet (I have only noticed a classifier for T1 images), so I am trying to use a custom classifier (using mriqc_clf) for the func images.

The preprocessed connectomes project (PCP) has QC metrics and ratings from human raters for the ABIDE and CoRR functional datasets, and I am trying to use these as the training datasets.

Not all IQMs between MRIQC and PCP overlap (aor, aqi, dvars_std, efc, fd_mean, fd_num, fd_perc, fwhm_avg). fber and gsr also overlap, however, the values look quite different.

How is fber calculated by MRIQC? For one of my datasets of N=~200 images, fber (as calculated by MRIQC) is 27392.78+/-7687.18 (M+/-SD) and another dataset, it is 598.05+/-118.79, whereas fber is 100.55+/-43.55 for ABIDE and 94.48+/-35.50 for CoRR (as calculated by PCP). I am wondering if there’s differences in how the fber metric is calculated?

Is it recommended that I should demean each variable (i.e. subtract the grand mean from each value) in both the training and testing datasets before classification using mriqc_clf?

2

We tried a little exploration of that in our poster for OHBM 2017:

OHBM2017-poster.jpg3370×2384 2.03 MB

Regarding the different implementations of FBER, you can compare in their repos:

Without looking at them, I’m fairly sure that the largest difference is on the input masks, in this case.

Is it recommended that I should demean each variable (i.e. subtract the grand mean from each value) in both the training and testing datasets before classification using mriqc_clf?

That depends on your classifier. SVMs will definitely need this. With random forests, we found that only site-wise scaling by the standard deviation was necessary. Please note that “site-wise”. Normalization was done per site. This is an information you don’t usually have at regression time.