docs/PAPER_EXPERIMENT_PLAN.md

# ReMAP-PET Paper and Experiment Plan

## Working Title

**ReMAP-PET: Region-guided Metabolic Alignment with Partial-tuned PET Encoders**

## Central Claim

FDG-PET should not be treated only as an image classification input. Its paired 120-region SUVR profile provides structured metabolic semantics. ReMAP-PET uses this region-level supervision to partially adapt a pretrained 3D medical/brain encoder into a PET metabolic encoder.

The core claim is:

> High-level partial tuning plus regional metabolic alignment produces a stronger FDG-PET representation than frozen probing, and it gives a usable bridge from 3D PET to structured metabolic language.

## Paper Structure

### 1. Introduction

Motivation:

- FDG-PET reflects glucose metabolism rather than anatomy.
- Existing brain MRI foundation models are not directly optimized for PET metabolic semantics.
- Current PET work often focuses on diagnosis labels, while our data provides richer region-level SUVR supervision.

Main contribution:

- Define PET-to-regional metabolic alignment.
- Propose ReMAP-PET, a partial-tuned PET encoder trained with SUVR regression and PET-SUVR contrastive alignment.
- Evaluate against frozen pretrained encoders and newer brain/neuroimaging foundation encoders.
- Show interpretable region-level prediction and retrieval.

### 2. Related Work

Groups to cover:

- Brain MRI foundation models: BrainIAC, BrainFound, BrainDINO.
- General clinical neuroimaging foundation models: NeuroVFM.
- 3D MRI vision-language models: Decipher-MR.
- Generative multimodal brain models: GenBrain.
- FDG-PET and AD metabolic imaging.
- Vision-language / CLIP-style contrastive alignment.

### 3. Dataset and Task

Dataset:

- 1015 paired FDG-PET volumes and SUVR CSV tables.
- 120 real brain regions after excluding `Background`.
- Fixed split: 710 train, 152 val, 153 test.

Tasks:

- PET-to-SUVR regression.
- PET-SUVR bidirectional retrieval.
- High/low metabolic region identification.
- Controlled region-text alignment in the later stage.

### 4. Method

ReMAP-PET Stage 1:

```text
3D FDG-PET
  -> pretrained 3D encoder
  -> partial tuning of high-level block
  -> PET embedding
  -> SUVR regression head
  -> contrastive alignment with SUVR embedding
```

Training loss:

```text
L = lambda_reg * L_suvr
  + lambda_con * L_pet_suvr_contrastive
```

Main implementation:

- PET encoder: MedicalNet 3D ResNet-50.
- Partial tuning: only `layer4`.
- Current best: `lambda_reg=1.0`, `lambda_con=0.2`.

### 5. Experiments

#### Table 1: Dataset Summary

Purpose:

- Show what data we have and why the task is not ordinary classification.

Columns:

- Dataset
- Samples
- PET format
- SUVR regions
- Train / Val / Test
- Labels available
- Task supported

Rows:

- Our FDG-PET-SUVR dataset.
- External ADNI/OASIS-3 if acquired later.

#### Table 2: Main Stage 1 Internal Test Results

Purpose:

- Main quantitative result on the held-out test set.

Rows:

- MedicalNet frozen MLP.
- BrainIAC frozen MLP.
- ReMAP-PET / MedicalNet layer4 partial tuning.
- BrainFound.
- NeuroVFM encoder.
- Decipher-MR.
- BrainDINO.
- GenBrain encoder.

Metrics:

- MAE
- RMSE
- Pearson
- Spearman
- PET->SUVR R@1 / R@5
- SUVR->PET R@1 / R@5
- MRR
- Median Rank
- Top5 High
- Top5 Low

#### Table 3: Modern Encoder Comparison

Purpose:

- Show whether newer brain/neuroimaging foundation encoders transfer better than the current ReMAP-PET backbone.

Rows:

- BrainFound.
- NeuroVFM encoder.
- Decipher-MR.
- BrainDINO.
- GenBrain encoder.
- ReMAP-PET as reference.

Settings:

- Frozen probing.
- Adapter tuning if feasible.
- High-level partial tuning if the architecture supports it.

Primary metrics:

- MAE
- Pearson
- PET->SUVR R@1
- SUVR->PET R@1
- Top5 Low

#### Table 4: ReMAP-PET Ablation

Purpose:

- Justify the ReMAP-PET training objective and show why the chosen partial-tuned alignment design works.

Rows:

- Regression only: `lambda_con=0.0`, `lambda_reg=1.0`.
- Weak contrastive: `lambda_con=0.1`, `lambda_reg=1.0`.
- Current setting: `lambda_con=0.2`, `lambda_reg=1.0`.
- Strong contrastive: `lambda_con=0.5`, `lambda_reg=1.0`.
- Contrastive-heavy: `lambda_con=1.0`, `lambda_reg=1.0`.

Metrics:

- MAE
- RMSE
- Pearson
- Spearman
- PET->SUVR R@1
- SUVR->PET R@1
- Top5 High
- Top5 Low

Expected interpretation:

- Regression-only gives the best numeric SUVR prediction but fails at PET-SUVR retrieval.
- Stronger contrastive weights greatly improve retrieval but gradually hurt numeric SUVR prediction.
- The balanced ReMAP-PET setting should be selected by the intended Stage 1 target: metabolic prediction plus cross-modal alignment, not regression alone.

Current test results are stored in:

```text
docs/REMAP_PET_ABLATION_RESULTS.csv
```

Main observation:

```text
lambda_con=0.0 -> MAE 0.0549, PET->SUVR R@1 0.0065
lambda_con=0.2 -> MAE 0.0700, PET->SUVR R@1 0.7778
lambda_con=0.5 -> MAE 0.0766, PET->SUVR R@1 0.8366
```

This means pure regression is not enough for the paper goal. ReMAP-PET needs contrastive alignment because the paper claims a PET-to-regional-metabolic semantic space, not only SUVR curve fitting.

#### Table 5: External Generalization

Purpose:

- Show the learned PET metabolic encoder is not only fitting our internal dataset.

Candidate external datasets:

- ADNI FDG-PET with PET Core SUVR summaries.
- OASIS-3 FDG-PET with post-processed PET outputs.

Rows:

- Frozen MedicalNet.
- ReMAP-PET.
- Best modern encoder variant.

Metrics:

- If SUVR available: MAE, RMSE, Pearson, Spearman.
- If diagnosis labels available: linear probe AUROC / balanced accuracy.
- Retrieval if paired PET-SUVR is available.

#### Table 6: Region Interpretability

Purpose:

- Demonstrate clinical interpretability at the brain-region level.

Rows:

- Model variants.

Metrics:

- Top5 Low overlap.
- Top5 High overlap.
- Region mention precision.
- Region mention recall.
- Example low-metabolism regions recovered.

### 6. Figures

#### Figure 1: Method Overview

Show:

```text
3D PET -> partial-tuned PET encoder -> PET embedding
SUVR table -> region encoder -> SUVR embedding
PET embedding <-> SUVR embedding contrastive alignment
PET embedding -> 120-region SUVR prediction
```

#### Figure 2: Embedding Retrieval Visualization

Show:

- PET-SUVR similarity matrix.
- Correct pairs on diagonal.
- Compare frozen MedicalNet vs ReMAP-PET.

#### Figure 3: Region-level Case Study

Show:

- Predicted vs true top low-metabolism regions.
- Controlled text generated from true/predicted SUVR.
- Example similar-case retrieval.

#### Figure 4: Ablation Tradeoff Curve

Show:

- x-axis: contrastive weight.
- y-axis 1: MAE.
- y-axis 2: PET->SUVR Recall@1.

## Experiment Execution Order

1. Finish ReMAP-PET loss-weight ablation.
2. Evaluate each best checkpoint on the internal test set.
3. Integrate the five chosen modern encoders:
   - BrainFound.
   - NeuroVFM encoder.
   - Decipher-MR.
   - BrainDINO.
   - GenBrain encoder.
4. Run frozen probing first for every modern encoder.
5. Add adapter or high-level partial tuning only where the architecture and weights are stable.
6. Add external validation on ADNI or OASIS-3.

## What Not To Claim

Do not claim:

- We solve clinical AD diagnosis directly.
- We train a full PET foundation model from scratch.
- We generate free-form radiology reports.

Safer claim:

> ReMAP-PET learns a region-grounded FDG-PET metabolic representation that supports SUVR prediction, PET-SUVR retrieval, and interpretable metabolic region analysis.