October 13, 2025Open Access

Data Scaling Laws for Radiology Foundation Models

View Full Paper

Ask AI

Bookmark

Continual pretraining of MI2 with structured supervision significantly enhances model performance.
Evaluation included diverse tasks: classification of radiology findings and segmentation of lines and tubes.
MI2 showed better scalability for finding-related tasks, while RAD-DINO excelled in tube-related tasks.
Surprisingly, as few as 30k in-domain samples can outperform existing foundation models in some evaluations.

Discussion

No takes yet. Share an insight, caveat, or question.

Systematic study demonstrates performance improvements in radiology tasks with continual pretraining, suggesting optimal data utilization.

Ilse et al. (2025) studied this question.

View Full Paper

Ask AI

Bookmark

Systematic study demonstrates performance improvements in radiology tasks with continual pretraining, suggesting optimal data utilization.