Main research findings
Gastrointestinal (GI) tract cancers are the leading cause of cancer-related death worldwide. Among GI cancers, colorectal cancer is the third most commonly diagnosed cancer. Colonoscopy is the gold standard for screening colorectal cancer, but it is an expensive, time-demanding, and operator-dependent procedure. Studies have reported polyp miss rates of 22%-28% during these procedures. Computer-aided diagnosis (CAD) methods can help to highlight suspicious lesions on the screen and alert gastroenterologists in real-time, improving the clinical outcome irrespective of operator experience, potentially saving millions of lives.
- We have designed machine learning-based architectures for GI tract abnormality and findings classification. Our classification algorithms classify different GI endoscopy findings with an accuracy of 98.07%.
- The availability of public datasets was one of the significant challenges for the development of automated methods. We addressed this problem by collecting, curating, annotating, and publicly releasing several datasets, including the world’s largest publicly available GI endoscopy and video capsule endoscopy datasets.
- Our algorithms can identify and segment the potential presence of lesions during colonoscopy with an accuracy of more than 94.93% and at a real-time speed of 182.38 frames per second. Moreover, our algorithm can simultaneously identify multiple polyps, including flat and sessile polyps that are often overlooked by endoscopists during the colonoscopy examination.
- We also performed instrument segmentation to be able to detect which instruments are used during the examination. The developed method is able to segment different types of surgical instruments in real-time.
- As another aspect of our work, we also addressed the challenge of generalizability, meaning that our models can perform well on completely new data enabling for example that a model can be moved from one hospital to another. We demonstrated reliable performance and high generalizability compared to baseline algorithms.
- Due to the challenge of not having access to high-end hardware in the hospitals, we have developed lightweight architectures that can be integrated with low-end hardware devices.
- Our algorithms are not only designed for polyp segmentation and surgical instrument segmentation but can also be exploited for other medical or non-medical image segmentation tasks. All of our work and algorithms are open-sourced and received very well by the community.
- Professor Morten Goodwin, University of Agder (UiA)
- Professor Niall Murray, Athlone Institute of Technology, Ireland
- Professor Anne Håkansson, Arctic University of Norway (UiT)
- Associate Professor Håvard D. Johansen, Arctic University of Norway (UiT)
- Professor Pål Halvorsen, SimulaMet
- Professor Dag Johansen, Arctic University of Norway (UiT)
- Chief Research Scientist Michael A. Riegler, SimulaMet
Chair of defense
- Professor Cordian Riener, Pro-dean, Arctic University of Norway (UiT)