DEAL

Group robustness-aware MRI training pipeline for Alzheimer's disease (AD) prediction, combining 3D MRI with tabular clinical/demographic features via multiple fusion methods.

Overview

DEAL supports:

• Targets: binary AD classification (ad), MCI-to-AD transition (ad_transition), and time-to-AD Cox regression (ad_time_cox)
• MRI backbones: MONAI ResNet, EfficientNet, Vision Transformer (ViT-T/S/B), with optional 2D sagittal view
• Fusion methods: feature concatenation, FiLM, FiLM+demographics, DAFT, tabular concatenation (with/without linear projection), tabular attention, and tabular-only baselines
• Group Robustness: Group DRO, SUBG, and age-based classifier decoupling

Requirements

uv venv --python=3.10

source .venv/bin/activate

uv pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
uv pip install monai --no-deps
uv pip install numpy==1.26.4
uv pip install scikit-learn scikit-survival pandas torchio wandb

AD/CN pretrained weights

Weights of pretrained models trained with AD/CN clssification are provided in https://drive.google.com/drive/folders/1VgpcKJDJJza9II9LzZVwy7OgQ96m8zel?usp=sharing.

Quick start

Train (example: FiLM + demographics, AD transition, pretrained ResNet):

python main.py \
  --method film_demo \
  --film-only-last \
  --film-aux-net mlp \
  --film-aux-net-act silu \
  --tabular-input mmse cdrsb adas11 age_dummy gender educat faq \
  --target ad_transition \
  --mri-arch monai_resnet18 \
  --resolution 1mm \
  --pretrained \
  --normalize \
  --epochs 30 \
  --batch-size 4 \
  --lr 1e-5 \
  --mri-modelpath /path/to/pretrained_mri.pt \
  --save-model \
  --date my_experiment

Main options

Category	Arguments
Data	--data-dir, --resolution (1mm / 2mm), --target (ad / ad_transition / ad_time_cox), --normalize, --merge-train-val, --sagittal
Backbone	--mri-arch (efficientnet, vit_t, vit_s, vit_b, monai_resnet18), --resnet-backbone, --efficientnet-backbone
Method	--method (scratch, feature_concat, film, film_demo, daft, tab_concat, tab_concat_lin_proj, tab_attention, tabular_only)
FiLM/DAFT	--film-aux-net, --film-aux-net-act, --film-only-last, --tabular-input, --daft-input
Optimization	--epochs, --batch-size, --lr, --optimizer, --scheduler, --weight-decay
Group Robustness	--group-dro, --mri-subg, --decouple, --group-weight-lr
Evaluation	--test-set-id, --seed, --last-eval-only, --no-best-th, --save-model

See argument.py for the full list.

Scripts

• scripts/run_DEAL.sh – example runs for DEAL
• scripts/run_deal_arch_abl.sh – architecture ablations (EfficientNet, ViT-B, 2D sagittal) for ad_transition
• scripts/run_baselines.sh – baseline configurations

Outputs

• Checkpoints: ./trained_models/<date>/mri/<target>/
• Results/logs: ./results/<date>/mri/<target>/

License

See repository for license terms.