samjohnson11/oasis-1-brainiac-dementia

OASIS-1 dementia classification: fine-tuned BrainIAC (CDR 0 vs CDR > 0)

Code: https://github.com/samwjohnson/oasis-1-dementia-classification

Status of weights

The ft235_nov2 run saved fine-tuned backbone and head weights: 25 per-fold checkpoints, 177 MB each in bf16, about 4.4 GB in total. They are not uploaded here yet. The checkpoints are in the author's private storage and will be added in a later update.

For now, this repo holds the run configs, aggregate results, and statistical-test outputs. If the weights are added later, they are derivatives of BrainIAC and fall under the BrainIAC Research-Only License.

Approach

• Backbone. BrainIAC ViT-B (MONAI ViT, 96³ input, 16³ patches, 768-d, no CLS token; the head reads token 0). It is initialized from the official BrainIAC.ckpt (Tak et al., 2026).
• Head. Dropout(0.2) + Linear(768→1), trained with BCEWithLogits.
• Optimizer. AdamW, lr 1e-5 for the backbone and 1e-3 for the head, weight decay 0.05. Grad clip 1.0, bf16 on a Colab L4. Up to 40 epochs with patience 8 and a 15% inner early-stopping split. Augmentation is a random L-R flip.
• Data. OASIS-1 cross-sectional, 235 subjects (135 CDR 0 / 100 CDR > 0).
• Evaluation. 5×5 stratified, subject-level nested CV, giving 25 held-out folds. The metric is mean AUROC over repeats, with its SD.
• Preprocessing variant (n4rigid). A-P flip, N4 bias correction in the brain mask, 1 mm resampling, COM initialisation, rigid Mattes-MI registration to the HD-BET head template, 96³ crop, in-mask z-score.

Results

All values are read from results/*/summary*.json.

Run	Mean AUROC (SD)
Frozen backbone, 198 subj	0.619 (0.028)
Full FT, 198 subj	0.643 (0.013)
Frozen backbone, 235 subj	0.716 (0.005)
Full FT, 235 subj (ft235b; Nov 2 rerun identical)	0.738 (0.003)
Last block unfrozen	0.730 (0.005)
Last 2 blocks, lr 1e-6 to 1e-5	0.729 to 0.730
Last 2 blocks, lr 1e-4	0.728 (0.009)
Full FT on N4 + rigid volumes	0.748 (0.007)

• Full FT vs frozen (235 subjects). +0.023 AUROC, better in 20 of 25 folds. The Nadeau–Bengio corrected p is 0.067.
• n4rigid vs ft235b. +0.009 AUROC, with corrected p = 0.45.
• Attention maps. These cover 4 test scans from one fold. In the deeper blocks, attention from token 0 concentrates toward the centre of the brain. In the last blocks, the 2 CDR > 0 scans showed stronger central concentration than the 2 CDR 0 scans. This is qualitative, and n is tiny.
• Paper reference. Tak et al. report AUC 0.88 (95% CI 0.75–0.97) for OASIS-1 MCI vs healthy control. That figure comes from a single 40-scan test set and is not directly comparable.

Limitations

• The cohort is small (235 subjects). The fold-level AUROC SD is about 0.065.
• The model was built for research only. It is not validated for any clinical use, and the BrainIAC license prohibits clinical or commercial use.
• Labels come from a CDR threshold, and CDR 0.5 is grouped with dementia.
• The model was trained on a single dataset from a single site, and there was no external test.

Data and licences

• No OASIS images or image-derived volumes or features are included here. Get OASIS-1 from https://sites.wustl.edu/oasisbrains/, which is for non-commercial research use only.
• Acknowledgement required by OASIS: Data were provided by OASIS-1: Cross-Sectional: Principal Investigators: D. Marcus, R. Buckner, J. Csernansky, J. Morris; P50 AG05681, P01 AG03991, P01 AG026276, R01 AG021910, P20 MH071616, U24 RR021382. Marcus et al. 2007, https://doi.org/10.1162/jocn.2007.19.9.1498
• BrainIAC: Tak D. et al., A generalizable foundation model for analysis of human brain MRI, Nature Neuroscience (2026), https://doi.org/10.1038/s41593-026-02202-6. Code and checkpoint: https://github.com/AIM-KannLab/BrainIAC