Researchers demonstrate that different 3D medical imaging domains (CT, MRI, PET) transfer knowledge asymmetrically during pretraining, following predictable power-law patterns. By optimizing data allocation based on these transfer dynamics, they achieve up to 58% performance gains over proportional sampling, revealing a hub-and-island structure where certain domains act as foundational knowledge sources for others.
This research addresses a fundamental challenge in building unified foundation models for medical imaging: how to efficiently allocate training data across heterogeneous domains when transfer between them is inconsistent and directional. The discovery of asymmetric knowledge transfer—where training on one domain significantly improves another but not vice versa—has substantial implications for model development efficiency and resource allocation in healthcare AI.
The work builds on the expanding interest in vision foundation models that can handle volumetric data, a critical need as medical institutions increasingly digitize imaging across multiple modalities. Previous approaches relied on heuristic mixing strategies without principled understanding of domain interactions. This research transforms that through quantifiable scaling laws, revealing that some imaging modalities function as knowledge hubs while others remain isolated islands requiring direct computational investment.
The practical impact is significant: hospitals and research institutions developing multimodal imaging systems can now strategically prioritize pretraining data rather than using proportional sampling. A 58% performance improvement with transfer-aware allocation translates to either faster model convergence or better downstream performance in disease classification and segmentation tasks—both critical for clinical adoption. The model's generalization to unseen budgets (r=0.989) suggests these principles remain robust across different resource constraints.
Looking forward, this framework potentially extends beyond medical imaging to other multimodal domains where transfer asymmetry exists. The hub-and-island structure provides a conceptual toolkit for understanding which data modalities deserve strategic investment in foundation model development, informing both research priorities and computational infrastructure decisions.
- →Knowledge transfer between medical imaging domains follows asymmetric, predictable power-law patterns rather than occurring uniformly
- →Transfer-aware data allocation achieves 58% performance improvement over proportional sampling in 3D medical imaging pretraining
- →Certain imaging modalities emerge as 'hub' domains that improve many others, warranting preferential computational allocation
- →The derived allocation strategy generalizes reliably to unseen budgets with 0.989 correlation, enabling practical application across resource constraints
- →Downstream validation confirms transfer-aware pretraining produces stronger representations for clinical disease classification and organ segmentation tasks