Visual Intelligence Online Seminar #99: A journey to finding the best radiological anomaly localization model

Abstract

Medical anomaly detection methods have long promised to revolutionize clinical workflows by supporting radiologists and highlighting overlooked findings. However, deploying these methods requires navigating unique domain-specific complexities, such as the needle-in-a-haystack nature of 3D volumes, detecting rare pathologies, and ensuring reliability in critical acute cases. This talk offers an overview of anomaly detection in radiological imaging and unpacks why current evaluation paradigms often fall short in this setting. I will present a framework for building benchmarks modeled after real-world clinical demands and map out the critical frontiers the computer vision community must conquer next.