Scaling law for 3D medical imaging boosts AI training efficiency by 58%

A new paper from Ho Hin Lee and collaborators at multiple universities introduces knowledge transfer scaling laws for 3D medical imaging. The team observed that when pretraining vision foundation models on mixed modalities (CT, MRI, PET), different domains scale at variable rates and knowledge transfer is strongly asymmetric: training on one domain (e.g., CT) can substantially improve performance on another (e.g., MRI), but the reverse may be much weaker. Both the MAE reconstruction loss and cross-domain transfer follow predictable power-law trends with domain-specific coefficients.

Using these insights, the researchers formulated data allocation as a scaling-law optimization problem. The derived allocations revealed an interpretable hub-and-island structure: highly transferable domains like CT emerge as hubs that strategically benefit many others, while isolated domains like PET act as islands requiring direct investment. Empirically, their transfer-aware allocation outperformed standard data-proportional sampling by up to 58% on downstream tasks including disease classification and organ/lesion segmentation. The method also generalized to unseen compute budgets with a correlation of r=0.989, suggesting robust practical value for building 3D medical foundation models.