Differential Privacy in Medical Imaging: Pretraining Domain Key to Utility

A recent study published on arXiv (2601.19618) explores the impact of pretraining domain and objective on the privacy-utility balance in differentially private medical image analysis. Researchers employed ConvNeXt classifiers trained with differentially private stochastic gradient descent, utilizing five varying initializations for domain and objective independently, across four privacy budgets and a no-privacy condition. They assessed the models on over 590,000 chest radiographs sourced from five external datasets spanning four countries. Notably, supervised pretraining on chest radiographs achieved the top rank in 24 out of 25 dataset and budget combinations, surpassing ImageNet initialization. The results indicate that the pretraining domain is more crucial than the objective for maintaining utility under differential privacy, suggesting that domain-specific pretraining data can enhance accuracy while ensuring privacy.