docs/EXPERIMENT_STATUS.md

Experiment Status

Current Method: ReMAP-PET

We use the name ReMAP-PET for the current Stage 1 method:

Region-guided Metabolic Alignment with Partial-tuned PET encoders

The method trains a PET encoder by aligning 3D FDG-PET volumes with 120-region SUVR profiles. The key design is partial tuning of the high-level encoder block rather than treating the pretrained encoder as a fixed feature extractor.

The best current route is:

MedicalNet 3D ResNet-50
+ layer4 partial tuning
+ PET-SUVR regression and contrastive alignment

Checkpoint on server:

/data/Albus/Brain/runs/foundation/medicalnet_layer4_regalign_best.pt

Training command:

CUDA_VISIBLE_DEVICES=0 /data/Albus/miniconda3/bin/python -u scripts/train_pet_foundation.py \
  --backbone medicalnet \
  --encoder-train-scope layer4 \
  --epochs 50 \
  --batch-size 4 \
  --lr 1e-5 \
  --num-workers 2 \
  --output-size 96 96 96 \
  --contrastive-weight 0.2 \
  --regression-weight 1.0 \
  --out runs/foundation/medicalnet_layer4_regalign.pt \
  --best-out runs/foundation/medicalnet_layer4_regalign_best.pt

Test Metrics

Model	MAE	RMSE	Pearson	Spearman	PET->SUVR R@1	SUVR->PET R@1	Top5 High	Top5 Low
MedicalNet frozen MLP	0.1173	0.1530	0.7390	0.8454	0.0261	0.0458	0.4288	0.7294
BrainIAC frozen MLP	0.1248	0.1614	0.7069	0.8150	0.0131	0.0196	0.3425	0.6797
ReMAP-PET (MedicalNet layer4 partial tuning)	0.0700	0.0894	0.9198	0.9350	0.7778	0.9216	0.6366	0.7660

Interpretation

The current result supports partial tuning of the final MedicalNet stage as the primary Stage 1 strategy:

• it preserves MedicalNet's pretrained 3D medical representation;
• it adapts high-level features to FDG-PET metabolism;
• it improves both SUVR regression and PET-SUVR retrieval;
• it uses little GPU memory.

Next Experiments

1. Add a region-token encoder instead of a flat SUVR MLP.
2. Add controlled text generated from SUVR tables.
3. Train PET-to-region-text contrastive alignment.
4. Add a loss-weight ablation:
   • regression only;
   • contrastive only;
   • contrastive 0.2 + regression 1.0;
   • contrastive 1.0 + regression 1.0.
5. Integrate BrainFM as a second foundation-model comparison.