Machine Learning in Medical Imaging: From Few Annotations to Heterogeneous Data and Interpretability
Henning Muller
May 12, 2021, Wednesday, 3:00 PM - 4:00 PM EDT
Abstract
The presentation highlights a few of the challenges encountered when building decision support tools using machine learning with medical images. Applications described are in the histopathology field as well as in radiology. Limited annotations are a general problem because particularly deep learning requires large amounts of data to obtain very good results. Using weak labels and models such as teacher/student approaches can help with this. To get acceptance for the tools by physicians interpretability/explainability of the outcomes is important. I will describe approaches beyond simple heat maps such as regression concept vectors that can help describe algorithmic decisions and allow to explore new biomarkers that can be understood by physicians.
Bio
Henning Muller is professor in computer science at HES-SO and in radiology at the University of Geneva. He studied medical informatics at the University of Heidelberg, Germany, then worked at Daimler-Benz research in Portland, OR, USA. From 1998-2002 he worked on his PhD degree at the University of Geneva, Switzerland with a research stay at Monash University, Melbourne, Australia in 2001. Since 2002 Henning has been working in medical informatics at the University Hospitals of Geneva where he habilitated in 2008 and was named titular professor in 2014. Since 2007 he has been a professor in business informatics at the HES-SO Valais in Sierre and since 2011 he has been responsible for the eHealth unit in Sierre. Henning was coordinator of the Khresmoi project, scientific coordinator of the VISCERAL project, initiator of the ImageCLEF benchmark. He has authored over 400 scientific papers, is in the editorial board of several journals and reviews for many journals and funding agencies around the world.