Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness

Researchers introduce Pan-FM, a foundation model trained on multimodal medical imaging from seven organs that addresses the critical problem of missing data in real-world biomedical datasets. The model uses Saliency-Guided Masking to prevent bias toward dominant organs and demonstrates superior performance on disease prediction tasks across the UK Biobank.

Pan-FM represents a meaningful advancement in medical AI by tackling a fundamental challenge in real-world healthcare data: missingness. Most existing foundation models train on isolated, single-organ datasets, whereas human health depends on coordinated biological processes across multiple organ systems. This research acknowledges that practical biomedical data rarely arrives complete—some patients lack certain imaging modalities—yet most AI systems fail gracefully under these conditions.

The innovation of Saliency-Guided Masking directly addresses shortcut learning, where models exploit dominant features rather than learning robust representations. By adaptively masking organs that the model relies on excessively, the approach encourages balanced cross-organ learning without significant computational cost. This is particularly valuable because it can integrate into existing self-supervised frameworks without requiring architectural redesign.

For the medical AI sector, Pan-FM's demonstration of improved robustness under missing-organ scenarios has substantial implications. Healthcare systems frequently encounter incomplete data due to equipment availability, cost constraints, or clinical protocols. A model that maintains predictive power despite missing modalities directly translates to broader real-world deployment potential. The validation on 13 disease categories and 14 single disease entities establishes clinical relevance beyond academic benchmarks.

The work signals movement toward more generalizable whole-body foundation models rather than isolated organ-specific systems. As healthcare organizations increasingly adopt AI, models that handle data heterogeneity gracefully will gain competitive advantage. Future development likely involves scaling to additional organs and exploring how whole-body representations improve transfer learning across different patient populations and healthcare systems.

• →Pan-FM successfully handles multimodal medical imaging from seven organs while managing realistic missing-data scenarios during training and inference
• →Saliency-Guided Masking prevents dominant-organ bias by adaptively masking organs the model over-relies on, improving balanced representation learning
• →The model outperforms single-organ and multi-organ baselines across 13 disease categories on UK Biobank data with enhanced robustness under missing modalities
• →The approach adds negligible computational overhead and integrates seamlessly into existing self-supervised learning frameworks
• →Robust handling of missing data addresses a critical gap between academic datasets and real-world healthcare environments