/ overview
A four-class Alzheimer's staging model built on an interpretable Multi-IT Transformer architecture with Triplet Attention, trained over roughly 44,000 MRI scans.
The emphasis was interpretability as much as accuracy: attention maps highlight the regions driving each prediction, so a clinician can inspect why a scan was placed in a stage.
The transformer was benchmarked head-to-head against four CNN baselines under identical preprocessing and augmentation, with training run on an NVIDIA A100.
/ what it does
- 01Four-class staging: non, very mild, mild and moderate demented.
- 02Triplet Attention for cross-dimension feature interaction with low overhead.
- 03Attention-map visualisation for per-prediction interpretability.
- 04Controlled comparison against four CNN baselines on the same splits.
- 05Mixed-precision training on an NVIDIA A100 to keep experiment cycles short.
/ numbers
44,000
MRI scans
4
classes
4 CNNs
baselines compared
/ stack
Modelling
PyTorchVision TransformersTriplet Attention
Data
MRI datasetsAugmentationStratified splits
Compute
NVIDIA A100Mixed precision