Knowledge Transfer Scaling Laws for 3D Medical Imaging

A recent investigation published on arXiv (2605.06859) explores the scaling laws associated with knowledge transfer in three-dimensional medical imaging. The study reveals that various imaging modalities—CT, MRI, and PET—exhibit differing scaling rates during the pretraining phase, characterized by asymmetric knowledge transfer: enhancing one domain through training can significantly benefit another, whereas the opposite effect is often less pronounced. Both the MAE reconstruction loss and the cross-domain transfer demonstrate consistent power-law patterns. The researchers frame data allocation as a scaling-law optimization challenge, uncovering a clear hub-and-spoke configuration. This research offers a systematic methodology for integrating diverse imaging domains within vision foundation models in the field of medical imaging.