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

2026-05-08 · Source: Takara TLDR - Daily AI Papers · Field: Science & Research — Artificial Intelligence & Machine Learning, Health & Medical Research, Life Sciences & Biology · Depth: Expert, quick

Summary

Pan-FM is a novel pan-organ foundation model designed for medical imaging, pre-trained on imaging data from seven organs: Brain, Heart, Adipose, Liver, Kidney, Spleen, and Pancreas. Unlike most existing foundation models that focus on unimodal data, Pan-FM addresses the challenge of realistic missing-organ scenarios in multimodal biomedical data. It employs a unified backbone to manage organ missingness during both training and inference, utilizing masking-based self-distillation. The model introduces Saliency-Guided Masking (SGM) to counteract "dominant-organ shortcut learning bias," where models over-rely on organs like adipose and heart. SGM adaptively masks dominant organs during pre-training based on model attention, fostering balanced cross-organ learning with negligible computational overhead. Evaluated on the UK Biobank, Pan-FM demonstrates superior prediction across 13 disease categories and 14 single disease entities, showing enhanced robustness in settings with missing organs.

Key takeaway

For AI Scientists developing multimodal medical imaging models, Pan-FM offers a robust approach to handling real-world missing data. Its Saliency-Guided Masking technique effectively mitigates dominant-organ bias, leading to more balanced and generalizable whole-body representations. You should consider incorporating similar adaptive masking strategies to improve model performance and robustness in complex, incomplete datasets.

Key insights

Pan-FM is a pan-organ foundation model robust to missing data via saliency-guided masking.

Principles

• Whole-body representations improve disease prediction.
• Dominant-organ bias hinders balanced multimodal learning.

Method

Pan-FM uses a unified backbone for missing data, pre-trained with masking-based self-distillation. Saliency-Guided Masking (SGM) adaptively masks dominant organs using attention distribution to balance cross-organ learning.

In practice

• Integrate SGM into self-supervised frameworks.
• Apply Pan-FM to system neuroscience for multimodal learning.

Topics

• Pan-FM
• Foundation Models
• Multimodal Medical Imaging
• Saliency-Guided Masking
• Organ Missingness Robustness

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Unified Multimodal Model for Brain MRI Imputation and Understanding — UniBrain unifies MRI imputation and understanding, integrating multimodal data and self-alignment for robust medical reasoning.
• CrossPan: A Comprehensive Benchmark for Cross-Sequence Pancreas MRI Segmentation and Generalization — Cross-sequence domain shift is the primary barrier to robust pancreas MRI segmentation.
• Impact of Synthetic Lesional MR Images in Automated Focal Cortical Dysplasia Detection in Low-Data Scenarios — Synthetic FCD-MRIs generated by conditional networks can augment real data, reducing labeled data needs for automated detection.
• Learning Sparse Latent Predictive Foundation Model for Multimodal Neuroimaging — Neuro-JEPA is a sparse multimodal neuroimaging foundation model using a latent predictive objective and Mixture-of-Experts.
• Systematically decoding pathological morphologies and molecular profiles with unified multimodal embedding — Multi-Embed unifies morphological and molecular data for superior disease understanding and prediction.
• Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps — Explainable AI methods are crucial for validating deep learning models in medical diagnostics, revealing decision plausibility beyond performance metrics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.