The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) is one of the world's leading conferences in the field of computer vision. This year's edition took place in June 2025 in Nashville, Tennessee (USA) and was once again characterised by high-calibre research and technological innovation.
The Regensburg Medical Image Computing (ReMIC) laboratory at OTH Regensburg achieved an outstanding success: the publication ‘OpenMIBOOD: Open Medical Imaging Benchmarks for Out-Of-Distribution Detection’, authored by Max Gutbrod, PhD student at ReMIC, and co-authors David Rauber, Danilo Weber Nunes and Christoph Palm, was accepted as a poster for the main conference programme. In view of over 13,000 submissions and an acceptance rate of 22.1 %, this already represents a remarkable success. With a review score of 4.66 out of 5, the paper is also one of the top-rated contributions at CVPR 2025. The paper presented by Max Gutbrod met with keen interest from the specialist community.
OpenMIBOOD is dedicated to a central topic for the safe use of artificial intelligence in medicine: the reliable recognition of out-of-distribution (OOD) data - i.e. image material that does not correspond to the training data of a model and can potentially lead to incorrect decisions. The benchmark covers 14 medical datasets from three specialities and compares 24 existing OOD recognition methods. A key finding of the work: results from OOD tests with natural images cannot be easily transferred to medical data - which underlines the need for specialised, domain-specific benchmarks.
The presentation in Nashville was made possible by the support of the Regensburg Center for Health Sciences and Technology (RCHST): both the start-up funding for the research work and the travel expenses were supported by RCHST funds.
OTH Regensburg thus made an important contribution to the ongoing discussion about the reliability of AI systems in safety-critical fields of application such as medical image processing.
