Guidance requested for MRI selection strategy in ADNI (CN–MCI–AD classification workflow)

10 views

Skip to first unread message

Most Sonia Islam

unread,

Apr 23, 2026, 3:26:54 AM (4 days ago) Apr 23

to Alzheimer's Disease Neuroimaging Initiative (ADNI) Data

Hello everyone,

I am new to working with the ADNI dataset and have recently started a structural MRI–based classification project involving the three diagnostic groups: CN, MCI, and AD. I would appreciate guidance from experienced users regarding best practices for selecting imaging data and building a consistent pipeline.

I am currently trying to decide several important design choices before finalising my dataset construction:

First, regarding the type of MRI data, I am unsure whether I should use the original/raw images, the preprocessed images provided by ADNI, or apply my own preprocessing pipeline starting from raw data. Since ADNI includes multiple acquisition sites and scanners, I would like to understand which option typically provides better reproducibility and classification performance in baseline CN–MCI–AD studies. Second, regarding image field strength, I noticed that ADNI includes both 1.5T and 3T scans. Some documentation suggests that later ADNI phases increasingly adopted 3T MRI for newly enrolled participants. Should I restrict the dataset to a single field strength (e.g., only 3T) to avoid domain heterogeneity, or is it acceptable to combine 1.5T and 3T scans when building a classification dataset? Third, regarding diagnostic labels, I am currently planning to use baseline diagnosis only (CN vs MCI vs AD). The documentation indicates that baseline diagnosis is generally considered more reliable than screening diagnosis. Is using baseline-only labels the recommended strategy for a first-stage classification study or for my full work I used only baseline else others? Fourth, regarding MRI modality selection, I understand that ADNI structural MRI protocols emphasise 3D T1-weighted scans (e.g., MP-RAGE) for morphometric analysis across sites. For a standard deep learning classification pipeline, should I restrict the dataset to T1-weighted MRI only? Fifth, after performing advanced search queries in the ADNI portal, I observed that multiple MRI scans exist per subject at the same visit. In such cases, what is the recommended strategy for selecting a single scan per subject so that the dataset remains consistent and avoids mixed acquisition conditions? Sixth, regarding longitudinal versus cross-sectional setup, I am currently planning a baseline-only cross-sectional classification experiment. Would it be better to begin with baseline scans only and later extend the study to longitudinal modelling? Seventh, regarding 2D versus 3D input representation, I am unsure whether I should start with slice-based 2D CNN models or directly use full-volume 3D CNN architectures for the initial phase of experimentation. Eighth, regarding image description metadata, I would appreciate advice on how to select consistent acquisition descriptions (for example, MP-RAGE variants across vendors) so that scanner-level variability is minimised.

Finally, I would be grateful for recommendations on:

which preprocessing steps are essential before training,
whether skull stripping and spatial normalisation should be performed manually or using provided processed images,
and which baseline model architectures are typically recommended for an initial CN–MCI–AD classification study using ADNI MRI.

My objective is to construct a leakage-free subject-level structural MRI classification pipeline for CN–MCI–AD using deep learning.

Thank you very much for your time and support. I am looking forward to learning from the community and setting up a reliable and reproducible ADNI MRI classification pipeline.