BRIDGE

BRIDGE is a two-stage framework for Alzheimer's disease diagnosis from brain PET. It first learns a brain-to-gut latent transport from paired whole-body PET and then performs evidence-guided, reliability-gated virtual-gut fusion using brain-only PET. The public code supports CN, MCI, and AD classification while keeping the deployment input brain-only.

Overview

Stage 1 — Brain-to-Gut Latent Transport. Organ-specific 3D encoders map paired brain and gut PET into a shared latent space. A conditional vector field is trained with flow matching and an endpoint constraint to predict a virtual-gut representation from the brain representation.

Stage 2 — Evidence-Guided Virtual-Gut Fusion. A diagnostic encoder is modulated by a reviewed disease- evidence anchor bank. The frozen Stage 1 transport produces virtual-gut evidence from brain PET. A residual head and confidence gate add only reliable virtual-gut information to the cerebral diagnostic logits.

At inference, BRIDGE requires only a brain PET volume. The offline evidence workflow is separate from model training and never receives patient images or labels.

Repository layout:

bridge/                 model, loss, data, and utility code
scripts/                Stage 1 and Stage 2 entry points
configs/                full and CPU demonstration configurations
examples/               anonymous manifest and anchor schemas
tests/                  end-to-end tensor smoke test
tools/                   pre-release sensitive-information scanner
assets/architecture.jpg method overview

Installation

Python 3.10 or newer is recommended.

git clone https://github.com/Idea89560041/BRIDGE.git
cd BRIDGE
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

Run the CPU smoke test:

python -m pytest -q

Data preparation

BRIDGE expects preprocessed NIfTI volumes in a common analysis space. Create two CSV manifests following examples/paired_manifest.csv and examples/diagnosis_manifest.csv.

• The paired manifest contains subject_id, brain_path, and gut_path for Stage 1.
• The diagnosis manifest contains subject_id, brain_path, label, and a fixed outer fold for Stage 2.
• Labels are CN, MCI, or AD; subject identifiers must be pseudonymous and unique.
• Paths in the examples are placeholders. Do not commit real manifests or images.

Stage 2 also expects a reviewed PyTorch anchor artifact containing an embeddings tensor of shape [K, D]. The embedding dimension D must match latent_dim in the selected configuration.

Training and Evaluation

Minimal reproducibility check

The following creates synthetic NIfTI volumes and runs one CPU epoch for both stages. Synthetic outputs are written under ignored data/ and runs/ directories.

python scripts/make_demo_data.py
python scripts/train_transport.py \
  --manifest data/demo/paired.csv \
  --config configs/demo.yaml \
  --output runs/demo_transport
python scripts/train_diagnosis.py \
  --manifest data/demo/diagnosis.csv \
  --anchors data/demo/anchors.pt \
  --transport runs/demo_transport/transport.pt \
  --config configs/demo.yaml \
  --fold 1 \
  --output runs/demo_diagnosis

Full experiment

Train Stage 1 once using paired whole-body PET:

python scripts/train_transport.py \
  --manifest /path/to/paired_manifest.csv \
  --config configs/bridge.yaml \
  --output runs/transport

Then train and evaluate Stage 2 independently for each fixed outer fold:

for FOLD in 1 2 3 4 5; do
  python scripts/train_diagnosis.py \
    --manifest /path/to/diagnosis_manifest.csv \
    --anchors /path/to/reviewed_anchor_embeddings.pt \
    --transport runs/transport/transport.pt \
    --config configs/bridge.yaml \
    --fold ${FOLD} \
    --output runs/diagnosis
done

Each fold writes selected checkpoints and metrics.json. Use validation data for model selection and evaluate the outer test fold only after selection. Keep fold assignments fixed across all reported experiments. See RESULTS.md for the reporting structure and DATA_AND_PRIVACY.md for release requirements.

License

This code is released under the MIT License. Dataset licenses and clinical governance requirements remain with their respective providers and institutions.