ACR DSI Lab pilot site
Current developments in AI to improve patient care are being driven by research happening primarily at institutions with extensive informatics and data science resources — and primarily using single-institution patient data. Yet repeated studies have shown that these deep learning models do not generalize well across institutional differences, such as patient demographics, disease prevalence, scanners, and acquisition settings[2] [3]. Although AI algorithms are more effective when trained on a wide and diverse array of clinical data, sharing data outside an institution is difficult due to patient privacy concerns. With ACR AI-LAB, the intention is to share AI models between institutions leaving patient information to remain on-site at the originating institution. Emory is one of the 7 sites participating to pilot the infrastructure required to create such a federated environment.
For the first part of this pilot, we are testing breast density algorithm brittleness across 10 Emory sites performing screening mammograms. This project has subprojects
1. Work on improving the performance of the breast density algorithm to improve robustness across the various Emory sites
2. Generate test datasets using phantoms to provide a baseline for testing for algorithm robustness
3. Generate synthetic datasets for testing algorithm robustness
4. Explore federated learning, differential privacy, and split learning across the 7 sites running the ACR AI lab pilot