Continual learning in medical imaging: what happens if the world moves on?
Date: Friday, May 06, 2022
Time: 1:30 PM to 2:30 PM Note: all times are in the Eastern Time Zone
Event Type: Seminar
Host: Polina Golland, CSAIL MIT
Contact: Sheila Sharbetian, 617-324-6747, firstname.lastname@example.org
Speaker URL: None
TALK: Continual learning in medical imaging: what happens if the world moves on?
Medical imaging characteristics can change over time due to novel acquisition technology or scan protocols. These domain shifts lead to a deterioration of machine learning model prediction accuracy. In this talk I will discuss a method relying on pseudo-domains to detect domain shifts in a continuous stream of imaging data, and to adapt models accordingly. Adaptation needs to expand model capabilities to new domains, while at the same time mitigating forgetting old domains. In the second part of the talk I will focus on how we can optimally choose newly arriving imaging data for annotation in such a situation, assuming a limited annotation budget. The method capabilities and limitations will be illustrated with tasks such as cardiac segmentation, lung nodule detection, and brain age estimation.
Created by Sheila Sharbetian at Monday, May 02, 2022 at 10:29 AM.