| name | pan-fm-pan-organ-foundation |
| description | Pan-Organ Foundation Model (Pan-FM) for multimodal biomedical imaging with missing-organ robustness. Pre-trained on seven organs (Brain, Heart, Adipose, Liver, Kidney, Spleen, Pancreas) using Saliency-Guided Masking (SGM) to prevent dominant-organ shortcut learning bias. Introduces masking-based self-distillation for handling realistic missing-organ scenarios. Use when: (1) building multi-organ medical imaging foundation models, (2) handling missing modalities/organs in biomedical data, (3) designing saliency-guided masking strategies, (4) developing whole-body representation learning systems, (5) addressing dominant-organ shortcut learning in multimodal pre-training. Activation keywords: pan-organ foundation model, missing organ robustness, saliency-guided masking, multimodal biomedical imaging, dominant-organ shortcut learning, whole-body representation learning, UK Biobank multi-organ.
|
Pan-FM: Pan-Organ Foundation Model with Saliency-Guided Masking
arXiv:2605.07055 | Wu, McIlvain, Yu, & Wen (2026)
Core Problem
Medical foundation models are trained on single organs in isolation, but human aging and disease
arise from coordinated biological processes across organs. Real-world multimodal biomedical data
have organs missing not at random (MNAR), causing:
- Reduced statistical power
- Limited generalizability
- Systematic bias
Key Innovation: Saliency-Guided Masking (SGM)
Dominant-Organ Shortcut Learning
Naive multimodal pre-training causes models to over-rely on dominant organs (adipose, heart)
as shortcuts, ignoring subtler cross-organ signals.
SGM Mechanism
- Compute attention distribution across organs during forward pass
- Adaptively mask dominant organs based on attention scores
- Force balanced learning — model must learn from all organs, not just the strongest signals
- Negligible overhead — integrates seamlessly into existing self-supervised frameworks
def saliency_guided_masking(attention_weights, threshold=0.7, mask_prob=0.3):
"""Mask organs with disproportionately high attention."""
organ_attention = attention_weights.mean(dim=-1)
dominant_mask = organ_attention > threshold
random_mask = torch.rand_like(organ_attention) < mask_prob
return dominant_mask & random_mask
Architecture
| Component | Description |
|---|
| Unified backbone | Single architecture handling all 7 organs with organ-specific adapters |
| Masking-based self-distillation | Teacher-student framework with organ masking during pre-training |
| Missing-organ handling | Built into training AND inference — no post-hoc imputation needed |
| Cross-organ attention | Models inter-organ dependencies explicitly |
Evaluation Results
- Dataset: UK Biobank
- Tasks: 13 disease categories + 14 single disease entities
- Baselines outperformed: Single-organ FMs and naive multi-organ models
- Key metric: Improved robustness under missing-organ settings
When to Use
Apply this skill when:
- Building foundation models spanning multiple organs/body systems
- Handling MNAR (Missing Not At Random) data in medical imaging
- Designing masking strategies for self-supervised learning
- Addressing modality imbalance in multimodal models
- Developing whole-body or system-level biomedical representations
Do NOT use when:
- Working with single-organ data only
- The task doesn't involve medical/biomedical imaging
- Missing data is MCAR (Missing Completely At Random) — simpler methods suffice
Implementation Patterns
Multi-Organ Pre-Training Pipeline
import torch
class PanFMPreTraining:
def __init__(self, organs, backbone, sgm_threshold=0.7):
self.organs = organs
self.backbone = backbone
self.sgm_threshold = sgm_threshold
def sgm_mask(self, attention_scores):
"""Saliency-Guided Masking to prevent dominant-organ shortcuts."""
organ_importance = attention_scores.mean(dim=-1)
dominant = organ_importance > self.sgm_threshold
mask = dominant & (torch.rand_like(organ_importance) < 0.3)
return mask
def forward_with_sgm(self, batch):
features, attention = self.backbone(batch)
mask = self.sgm_mask(attention)
teacher_features = features.detach()
student_features = self.backforward(batch, mask=mask)
return self.distillation_loss(teacher_features, student_features)
Missing-Organ Inference
def predict_with_missing_organs(model, available_organs, missing_mask):
"""Handle inference when some organs are missing."""
return model.predict(available_organs, mask=missing_mask)
Best Practices
- Always use SGM when pre-training on heterogeneous multi-organ data
- Monitor attention distributions — if one organ dominates >60% of attention, SGM is essential
- Test missing-organ robustness as part of model evaluation
- Report per-organ contribution to final predictions
- Use UK Biobank as benchmark for multi-organ model comparison
Related Frameworks
- Single-organ FMs: brain-only, heart-only models (baseline comparison)
- Naive multi-organ: concatenation-based approaches (prone to shortcut learning)
- Self-supervised learning: DINO, MAE frameworks (SGM integrates with these)
arXiv Reference
- Paper: "Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness"
- Authors: Qiangqiang Wu, Grace McIlvain, Zhou Yu, Junhao Wen
- ID: arXiv:2605.07055v1 | Categories: cs.CV, cs.AI | Date: 2026-05-08
- Key contribution: First foundation model pre-trained on 7 organs with SGM for MNAR robustness
Related Skills
- tribe-v2-foundation-model: Multi-modal brain foundation model (video/audio/language)
- brain-foundation-biomarker-validation: Foundation model biomarker validation
- brain-foundation-model-batch-effects: Batch effects in brain foundation models