Pan-FM: Foundation Model for Missing Organ Data

Pan-FM is an innovative foundation model created to address the issue of absent organ data in medical imaging. It has been pre-trained on seven organs: Brain, Heart, Adipose, Liver, Kidney, Spleen, and Pancreas. This model tackles the problem of multimodal biomedical data, which often has missing elements that are not randomly distributed, potentially diminishing power, constraining generalizability, and introducing bias. By employing a unified backbone and a masking-based self-distillation approach, Pan-FM effectively manages organ missingness during both training and inference. Researchers discovered that traditional multimodal pre-training can lead to shortcut learning bias favoring dominant organs, a challenge that Pan-FM seeks to address. Further details can be found in arXiv:2605.07055.