Latent Neurodegenerative Signatures
from Structural MRI
A publication-grade, OASIS-1–only pipeline combining anatomical features and clinical anchors to learn stable MRI representations of early neurodegenerative change. Results are presented for research purposes only.
Single-site, high-quality
CDR 0 vs 0.5. Late Fusion AUC 0.80.
Multi-site, 629 subjects
CN vs MCI/AD. Level-MAX: 0.81 AUC.
Robustness validated
OASIS↔ADNI transfer shows MRI stability.
Data quality matters
Single-site OASIS beats multi-site ADNI.
Zero leakage verified across all experiments
+16.5% with biomarkers (no cognitive scores)
+11% with temporal biomarker change
0.725 AUC alone - beats cognitive tests
Research Portal
View All →We observed a distinct direction-dependent robustness. MRI-Only models generalize better when transferring from OASIS to ADNI, while Late Fusion is superior in the reverse direction.
Quick Statistics
Comparison of fusion strategies for early dementia detection (without MMSE)
• MRI-only (ResNet18 512-dim embeddings) achieves AUC of 0.78 for CDR=0 vs CDR=0.5 classification.
• Late Fusion (MRI + Clinical) reaches AUC of 0.80 combining imaging and demographic features.
• Attention Fusion achieves AUC of 0.79 with learned modality weighting.
Key Research Findings
Why fusion models underperform
512 strong MRI features + 2 weak features (Age, Sex) = dimension imbalance
Clinical encoder creates 30 dims of noise, diluting MRI signal
Level-MAX: 0.808 AUC with CSF biomarkers (ABETA, TAU, PTAU) + APOE4 + Volumetrics
7 major cleaning steps applied
Subject-level de-duplication (1,825 to 629)
Baseline-only selection (no temporal leakage)
Subject-wise splits (zero overlap verified)
MMSE/CDR-SB excluded (no circular features)
Temporal biomarker analysis
Hippocampus: 0.725 AUC alone
+ Longitudinal Δ: 0.848 AUC
APOE4 carriers: 2x risk
Random Forest beats complex LSTM