Trial lecture
“Opening the Black Box: Explainable AI and Trust in Medical Diagnostics”
Time and place: June 3, 2026 11:15 AM. Kristen Nygaards sal (5370), Ole-Johan Dahls hus.
Main research findings
From simple image labels to useful medical video outlines
This thesis shows that artificial intelligence can outline important findings in medical videos without needing thousands of hand-drawn training masks. The work focuses on endoscopy, where doctors use a camera to examine the stomach and intestines, and where early detection of polyps can be important for preventing cancer.
Medical videos contain valuable information, but exact outlines of findings are slow and costly to create. This makes it hard to build reliable tools that can support clinicians. Therefore, the thesis explores a simpler path. It uses ordinary image labels, such as “polyp”, to guide powerful image models toward the right area in a video.
The thesis contributes the open HyperKvasir dataset, methods for automatic guidance of image models, and the Map2VidSeg video pipeline. The solution produced video outlines that were competitive with methods trained on detailed manual masks. This can reduce annotation work, speed up research, and make medical artificial intelligence easier to develop in fields where expert-labelled data is scarce.
Adjudication committee
- Professor Cheng-Hsin Hsu, Dept. of Computer Science, National Tsing Hua University, Taiwan
- Professor Anne Håkansson, Department of Computer Science, University of Tromsø, Norway
- Professor Xing Cai, Department of Informatics, University of Oslo, Norway
Supervisors
Associate Professor Håkon Kvale Stensland, Simula Research Laboratory/ Department of Informatics, University of Oslo, Norway
Professor Pål Halvorsen, Simula Metropolitan Center for Digital Engineering (SimulaMet), Norway
Professor Michael Riegler, Simula Research Laboratory, Norway
Chair of defence
Contact information at Department: Pernille Adine Nordby