Multi-Institution Encrypted Medical Imaging AI Validation Without Data Sharing

Adoption of artificial intelligence medical imaging applications is often impeded by barriers between healthcare systems and algorithm developers given that access to both private patient data commercial model IP important perform pre-deployment evaluation. This work investigates a framework for secure, privacy-preserving AI-enabled inference using CrypTFlow2, state-of-the-art end-to-end compiler allowing cryptographically secure 2-party Computation (2PC) protocols the machine learning vendor target owner. A common DenseNet-121 chest x-ray diagnosis was evaluated on multi-institutional radiographic datasets with without CrypTFlow2 two test sets spanning seven sites across US India, comprising 1,149 images. We measure comparative AUROC performance insecure in multiple pathology classification tasks, explore output distributional shifts resource constraints introduced inference. Secure demonstrated no significant difference all diagnoses, outputs from methods were distributionally equivalent. The use may allow off-the-shelf 2PC AI vendors imaging, changes performance, can facilitate scalable infrastructure real-world evaluation exposure or IP.