Please join us for a special seminar sponsored by the Department of Biostatistics and Medical Informatics and the Morgridge Institute for Research as part of the Morgridge/UW-Madison Biomedical Imaging Seminar Series.
Title: Strategies for learning representations of cellular morphology
Presenter: Dr. Juan Caicedo, Schmidt Fellow and Investigator, Broad Institute of MIT and Harvard University
Date: Thursday, December 15, 2022 from 1:30 to 2:30 PM
Location: H.F. DeLuca Forum, Discovery Building, 330 N. Orchard St, Madison, WI
Zoom Link: https://us02web.zoom.us/j/88189307843?pwd=WEdpdExWcjQ2VEhkTXdybHk4TFNYUT09
Zoom meeting ID: 881 8930 7843;
Zoom passcode: 916453
Abstract: Microscopy images are fundamental for biological research, and quantifying cellular phenotypes is at the core of numerous applications in drug discovery, functional genomics and personalized medicine. While deep learning has been successful to classify objects in natural images, its application in microscopy brings unique challenges for enabling reliable biological analysis. In this talk, two major challenges for learning representations of cellular morphology will be discussed: 1) lack of ground truth annotations for learning, 2) unwanted factors of variation that confound learning and downstream analysis. To tackle these challenges, I will present strategies based on weakly supervised and self-supervised learning, as well as approaches to correct confounding factors of variation to obtain robust representations of cellular phenotypes for biological analysis.