Adaptive Differential Privacy for Federated Medical Image Segmentation Across Diverse Modalities

Researchers Puja Saha and Eranga Ukwatta have introduced a novel framework called Adaptive Differentially Private Federated Learning (ADP-FL) designed to solve a critical bottleneck in medical AI. Federated learning allows hospitals to train models collaboratively without centralizing sensitive patient data, but adding standard differential privacy for stronger security often cripples model accuracy and training stability. ADP-FL tackles this by dynamically adjusting the amount of privacy-preserving noise injected during training based on real-time model performance, creating a better balance between data utility and privacy protection.

In rigorous testing across diverse medical imaging tasks—including skin lesion segmentation in dermoscopy, kidney tumor segmentation in 3D CT scans, and brain tumor segmentation in multi-parametric MRI—ADP-FL consistently outperformed conventional methods. It achieved significantly higher Dice scores, improved boundary delineation, faster convergence, and greater training stability, with performance approaching that of non-private federated learning. The paper, which won the CAD Best Paper Award at SPIE Medical Imaging 2026, demonstrates that high-accuracy, privacy-preserving AI for diagnostics is now more viable for real-world, multi-institutional use.