AI Insight
This study examined how different pretraining strategies affect the performance of medical imaging AI models when differential privacy protections are applied. Researchers trained chest X-ray classifiers using five different initialization methods across varying privacy levels and tested them on over 590,000 chest radiographs from multiple countries. They found that models pretrained on medical images (chest X-rays) substantially outperformed those pretrained on general images (ImageNet), with the performance gap widening as privacy protections increased, and that the domain of pretraining data mattered 2-3 times more than the training method used.
Why it matters
This research provides critical guidance for developing privacy-protected medical AI systems, demonstrating that using medical images for pretraining is essential for maintaining diagnostic accuracy under privacy constraints. The findings suggest that investing in domain-specific pretraining datasets, even when that pretraining must also be done privately, yields better results than relying on publicly available general-purpose image datasets.
Understand the Science
arXiv:2601.19618v2 Announce Type: replace-cross
Abstract: Differential privacy protects the patients whose images train medical imaging models, but it lowers diagnostic accuracy, and the initialization is the strongest known remedy. Practice increasingly favors large generic self-supervised encoders. Yet the pretraining objective and the pretraining domain are confounded in existing comparisons, so which one preserves utility under privacy is unknown, and the pretraining corpus is treated as public even when it holds patient images. We trained ConvNeXt classifiers with differentially private stochastic gradient descent from five initializations that vary the objective and the domain independently, at four privacy budgets and without privacy, and evaluated them locally on more than 590,000 chest radiographs from five external datasets in four countries. Supervised pretraining on chest radiographs ranked first in 24 of 25 dataset and budget combinations. Its lead over ImageNet grew from 2.5 to 14.6 points of macro-averaged area under the receiver operating characteristic curve as the budget tightened, and the domain effect exceeded the objective effect by a factor of 2.2 to 3.4. Pretraining that corpus privately cost about 5 points and, under privacy, still beat every public initialization. Low-rank adaptation removed about half the residual gap, and in-domain pretraining raised the worst-performing demographic subgroup. Under privacy, what a model was pretrained on outweighs how it was pretrained.