medarc/adni-mini-v1-2

ADNI Mini v1.2

Longitudinal-rich ADNI sMRI evaluation set for frozen-feature / linear-probe prognosis evaluation (clinical path, cognitive decline, AD-typical atrophy).

• 1,200 subjects / 1,200 scans (one processed T1w per subject)
• Soft diagnosis balance: 450 CN / 500 MCI / 250 AD
• Sex: 600 Female / 600 Male
• Age at scan: 60–90
• Single eval split as Hugging Face Arrow shards with embedded NIfTI bytes
• Brain masks are not included

Load

from datasets import load_dataset

ds = load_dataset("medarc/adni-mini-v1-2", split="eval")
print(ds)
# image column: nifti

Primary label coverage (approx.)

Target	n
ADAS-Cog13 annualized slope (48m)	~1121
Logical Memory delayed (LDELTOTAL) slope (48m)	~1031
Hippocampus / lateral ventricle / amygdala slopes (48m)	~1093 each
MCI → AD risk (known at 36m)	500 (186 events)
MCI A+ → AD (36m)	249 (101 events)
CN A+ → MCI/AD (36m)	134 (23 events)
Amyloid centiloid	~975
Tau meta-temporal SUVR	~533
CSF Aβ / p-tau / t-tau	~901 / 899 / 901

Task philosophy

Primary	Secondary
Staging (AD vs CN, CN/MCI/AD)	Molecular: centiloid, tau SUVR, CSF
MCI conversion risk 36m	A+ stratified progression
ADAS13 + LDEL slopes	Sex (sanity)
Hippo / lat. ventricle / amygdala slopes
Age, BAG, SynthSeg volumes

Orientation

• ADAS13 slope: higher = worse cognition
• LDELTOTAL slope: more negative = memory decline
• Hippocampus / amygdala slopes: more negative = atrophy (ICV-normalized)
• Lateral ventricle slope: more positive = expansion (L+R lateral only, not 3rd/4th)

Design notes

• Optimized for longitudinal richness (follow-up, multi-T1, conversion known), not pure 400/400/400 B-equal.
• Continuous cognitive/atrophy rates preferred over coarse ordinal scores (no CDR-SB primary).
• Molecular PET/CSF labels retained as secondary association checks (not “virtual PET” headlines).
• One scan per subject; longitudinal labels precomputed onto that baseline row.

Rebuild (local)

.venv/bin/python datasets/adni/select_v12.py \
  --output artifacts/adni-mini-v1-2/selected_manifest.csv

.venv/bin/python datasets/adni/build_mini.py \
  --selected-manifest artifacts/adni-mini-v1-2/selected_manifest.csv \
  --output-dir artifacts/adni-mini-v1-2 \
  --fingerprint adni-mini-v1-2-eval-1200

See datasets/adni/v1.2_notes.md for full design and feasibility notes.