Core Concepts
Privacy-preserving training of diagnostic deep learning models is feasible with excellent accuracy and fairness in real-life clinical datasets.
Abstract
Training large-scale AI models in medical imaging while preserving privacy is crucial. The study evaluates the impact of differential privacy (DP) on model accuracy and fairness using two datasets: clinical chest radiographs and 3D abdominal CT images for classifying pancreatic ductal adenocarcinoma. Despite lower accuracy, privacy-preserving training did not amplify discrimination based on age, sex, or co-morbidity. Differential privacy ensures protection against data reconstruction attacks and offers a formal framework for safeguarding individual data points. The study demonstrates that DP allows for high diagnostic accuracy and fairness in challenging real-world clinical scenarios.
Stats
N = 193 311 (large dataset of clinical chest radiographs)
N = 1 625 (dataset of 3D abdominal CT images)
AUROC (Area Under the Receiver-Operator Characteristic Curve)
Quotes
"Privacy-preserving training of diagnostic deep learning models is possible with excellent diagnostic accuracy and fairness."
"Differential privacy ensures protection against data reconstruction attacks and offers a formal framework for safeguarding individual data points."