Hi all,
I’ve just built a multimodal ADNI dataset for an Alzheimer’s-disease classification project and could use some advice on the fMRI side.
- Subjects: 172 total — 60 AD / 60 CN / 52 MCI
- T1-weighted: identical plain “MPRAGE” series for every subject (no GRAPPA, SENSE, etc.).
- rs-fMRI: single-band BOLD only (TR ≈ 3 s), but four different series labels appear:
Resting State fMRI
Axial rsfMRI (Eyes Open)
Axial fcMRI (EYES OPEN)
MoCoSeries
(ADNI’s on-the-fly motion-corrected export)
No multiband, no perfusion/ASL. Age difference between each subject’s MPRAGE and fMRI ≤ 1 year (usually 0).
My questions
- Are these four labels truly interchangeable once I run fMRIPrep?
- I know slice-timing and motion correction will be applied anyway, but does starting from
MoCoSeries
(already motion-corrected) introduce downstream bias or over-correction compared to the raw runs?
- Would you exclude any of these labels to minimise variability?
- For example, keep only
Resting State fMRI
andrsfMRI
and dropfcMRI
/MoCoSeries
, at the cost of losing ~15 % of subjects?
- If I keep them all, what is the best practice for harmonising before feeding a CNN?
- ComBat on extracted FC matrices?
- Add one-hot “label type” covariates?
- Simply rely on fMRIPrep’s outputs + data augmentation?
- Any gotchas with FreeSurfer
-3T
+ fMRIPrep when the T1s are uniform but fMRI labels differ?
The end goal is a multimodal deep-learning model (T1 + rs-fMRI) with XAI (Grad-CAM / SHAP) explanations. I’d love to hear what has worked (or failed) for others tackling similar label heterogeneity in ADNI.
Thanks in advance!