Main content start
AIMI Research Meeting: A Federated Learning Study on Data Heterogeneity and Privacy using Chest CT - Edward Lee, PhD
Thursday, February 2, 2023
3:00pm - 4:00pm PST
Hybrid: In-Person | Virtual
This event is open to:
Significant advances have been made in AI benefitting from the volume and diversity of data. However, AI systems for medicine reported in literature have had exposure to relatively less data from less diverse sources. This difference in scale can be largely attributed to the fact that gathering data from many hospitals to one central location is not only time-consuming but also poses data privacy risks. Federated learning (FL) is a field that enables learning across many hospitals without sharing private patient data. In this work, we demonstrate various FL methods on one of the largest and most diverse COVID-19 chest CT datasets touching all continents except Australia and Antarctica. Hospitals with over 10,000 patients and over 1 million images are collected from around the world. We show that FL techniques can achieve close to performance parity to Centralized Data Sharing (CDS) while also maintaining high performance across all sites with small, underrepresented data. We investigate the strengths and weaknesses for all approaches on this diverse dataset including the robustness to non-IID diversity of data. We briefly investigate measures of privacy.
Edward Lee is a postdoctoral fellow working on machine learning for healthcare with Prof. Kristen Yeom. He is interested in applied ML, AI infrastructure, and distributed optimization. He has co-developed logarithmic networks, which enables models to be compressed and communicated with 4-bits of precision with minimal loss in accuracy. He has received his Ph.D. in Electrical Engineering working on developing efficient computing hardware and algorithms for deep learning.
Attendance is open to the Stanford community. If you would like to attend in-person or on Zoom, please contact the AIMI Center at email@example.com.