Bayesian uncertainty estimation for detection of long-tail and unseen conditions in abdominal images
In person: 23 February 2022 • 11:30 AM - 11:50 AM PST
There is an unmet need for deep learning that could automatically adapt to the real world conditions of imbalanced medical imaging data. We applied uncertainty estimation to the representation learning of long-tailed and out-of-distribution samples. By estimating the relative uncertainties of the input data with a dynamic Monte-Carlo dropout and combination of losses, our Bayesian framework is able to adapt to the imbalanced data for learning generalizable classifiers. Our evaluation based on two public semantic segmentation datasets with different class imbalance ratios showed that the proposed framework generalizes to the different datasets better than existing state-of-the-art models.
Ludwig-Maximilians-Univ. München (Germany)
Mina Rezaei is a Postdoctoral Fellow and a Lecturer in the chair of Statistical Learning and Data Science, Department of Statistics at the Ludwig-Maximilians-University Munich (LMU). Previously, she earned her Ph.D. in deep learning methods for medical image analysis at Hasso-Plattner Institute (HPI), Potsdam University under the supervision of Prof. Dr. Christoph Meinel. She earned her M.Sc. degree in Artificial Intelligence at the Department of Computer Science, Shiraz University, and a Bachelor's degree in Computer Science, Software Engineering. Her research interests include both technical and theoretical skills of machine learning and health care applications.