When
Where
Title
Learning Ideal Observers and Observer-Aware Image Quality Metrics
Abstract
This talk presents two new ideas for advancing task-based image quality assessment using modern machine learning and information theory.
First, we will describe a new method for approximating the Bayesian ideal observer (IO) for binary signal-detection tasks. Although the IO defines the upper limit of performance and is widely used in principle to optimize medical imaging systems and acquisition strategies, its test statistic is analytically intractable in most practical settings. Recent deep learning approaches can approximate the IO, but they are largely black-box methods and do not ensure outputs with the mathematical structure of a likelihood ratio. To address this, we develop a data-driven approach based on the variational information bottleneck (VIB). The VIB learns a low-dimensional latent representation that preserves task-relevant information. By encouraging this latent representation to be multivariate Gaussian, an analytic likelihood-ratio test statistic can be computed in latent space.
Second, we will introduce new figures of merit for task-based image quality grounded in recent information-theoretic results. We propose predictive V-information as an objective, task-specific image-quality metric that accounts for a specified family of sub-ideal observers, making it an information-theoretic analogue of sub-ideal observer performance. In this sense, it quantifies not just how much task-relevant information is present in an image (or raw measurements), but how much of that information can be exploited by a constrained observer class when performing a downstream inference. We will discuss its relationship to the previously proposed task-specific information metric and illustrate its usefulness in a stylized magnetic resonance image restoration problem involving signal detection and discrimination with synthetic data.
Bio
Dr. Mark A. Anastasio is the Mallinckrodt Endowed Professor of Imaging Science and Vice Chair of Imaging Sciences and AI Research in the Mallinckrodt Institute of Radiology at Washington University in St. Louis. He is also the founding director of the Center for Computational and AI-enabled Imaging Sciences and a Professor of Electrical and Systems Engineering. Previously, he served as the Donald Biggar Willett Professor and Head of the Department of Bioengineering at UIUC. He is a Fellow of the IEEE, SPIE, American Institute for Medical and Biological Engineering (AIMBE) and International Academy of Medical and Biological Engineering (IAMBE). Dr. Anastasio’s research broadly addresses computational image science, tomographic image reconstruction, and the use of machine learning for applications in imaging science. He has made a wide range of contributions related to the computational aspects of wave-based imaging modalities that include optoacoustic/photoacoustic computed tomography and ultrasound computed tomography (UST). Dr. Anastasio has also been actively engaged in computational imaging science research related to the objective assessment of image quality and optimization of imaging system performance by use of modern machine learning methods. Earlier in his career, he was the recipient of a National Science Foundation (NSF) CAREER Award to develop image reconstruction methods for X-ray phase-contrast tomography. More recently was the recipient of the 2025 IEEE EMBS William J. Morlock Award for outstanding contributions to image reconstruction and biomedical imaging technology.
Can't Join Us In Person?
Register for the Zoom Webinar!
Subscribe to Upcoming Colloquium Announcements
Visit our website for future lecture dates and speaker information