File size: 2,610 Bytes
54c74d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
---
license: other
tags:
  - alzheimer
  - oasis-1
  - structural-mri
  - multimodal
  - dementia-staging
pretty_name: TriFuse-AD OASIS-1 Three-Stage Dementia Staging
---

# TriFuse-AD: Honest Multimodal Benchmark for Three-Stage Dementia Staging on OASIS-1

Code, processed data, results, and paper for a leakage-free benchmark of three-stage
cognitive classification (CN / VMD / AD) on the OASIS-1 cross-sectional cohort, plus
the proposed **TriFuse-AD** model (tri-planar CNN + slice-plane Transformer + gated
demographic fusion).

## Key result (honest / negative)

On an age-restricted cohort (≥60, 198 subjects) with subject-level repeated 5-fold CV
(3 seeds, 15 runs/model), **no MRI-only network beats a plain tabular XGBoost
(Macro-F1 0.474)**, and TriFuse-AD (0.488 ± 0.066) does **not** significantly beat a
trivial DenseNet late-concat baseline (0.497 ± 0.062; paired permutation p = 0.55). A
no-MRI structured model reaches Macro-F1 0.480 — most recoverable signal is
morphometric/demographic, not learned from raw voxels. No clinical / diagnostic /
SOTA / MCI / cross-site claims.

## Repository layout

| Path | Contents |
|------|----------|
| `src/` | `trifuse` package: data, models, training, eval, analysis |
| `scripts/` | experiment runner, table/figure/interpretability builders |
| `configs/` | model configs |
| `results/` | per-model OOF preds, summaries, tables, 27 figures |
| `paper/trifuse_ad.md` | full paper draft |
| `data_processed.zip` | preprocessed 2.5D + 3D arrays + `subjects_clean.csv` (1.5 GB) |
| `data/raw/*.tar.gz` | OASIS-1 cross-sectional discs 1–12 (16 GB) |

## Reproducing

```bash
pip install -r requirements  # torch cu128, timm, monai, nibabel, xgboost, sklearn, ...
unzip data_processed.zip                       # -> data/processed_2d, processed_3d, metadata
python scripts/run_experiments.py --grid main      # 11 models x 15 runs
python scripts/run_experiments.py --grid ablation  # 6 variants
python scripts/make_tables.py && python scripts/make_figures.py
```

## Cohort

OASIS-1, age≥60 → 198 subjects (CN=98, VMD=70, AD=30). Labels from CDR (0→CN,
0.5→VMD, ≥1→AD). CDR and MMSE are **never** model inputs (label leakage). One volume
per subject (`*_111_t88_masked_gfc`).

## License / data use

The `data/raw/` tarballs are the original **OASIS-1** cross-sectional release
(Marcus et al., 2007), redistributed here for reproducibility. OASIS data are subject
to the OASIS data-use terms; if you use them, cite the OASIS project and comply with
their agreement. Code and derived results in this repo are provided for research use.