New AI solution from VRVis boosts confidence in computer-aided diagnosis and receives the Austrian IT Award 'eAward 2021'.
Modern radiology is increasingly relying on computer-aided solutions to speed up hospital workflows. The direct comparison of different visualization methods for determining crucial image areas of an AI classification system: (a) the original with an annotated finding by a radiologist, (b) the VRVis method, (c) an industry standard solution (GradCAM).
Together with our long-standing partner Agfa Healthcare, VRVis has been working for two decades on optimizing workflows in digital radiology through state-of-the-art visual computing technologies. One of the results was awarded with the eAward 2021 as leading "top application of 'Trustworthy AI' from Austria". (c) Milena Krobath
Dimitrios Lenis and David Major (center), researchers in our Biomedical Image Informatics Group, received the eAward for the Interpretable Artificial Intelligence solution they developed for digital radiology. (c) Milena Krobath
The eAward showcases the best IT projects from Austria.
Every year, the eAward honors Austrian IT projects with a special business focus. This year, the VRVis research group Biomedical Image Informatics is honoured to receive the eAward in the category "Artificial Intelligence". The VRVis team's solution, which was developed in close collaboration with long-time company partner and global radiology solutions provider AGFA Healthcare, boosts confidence in computer-aided diagnosis thanks to the use of explainable artificial intelligence (XAI). The jury called the solution "Top application of 'Trustworthy AI' from Austria".
One of the key VRVis research topics is accelerating radiology workflows through the use of artificial intelligence. This includes methods that make the machine decision-making process explainable and comprehensible. This is an important prerequisite for increasing the trustworthiness of artificial intelligence solutions - which is particularly important for companies from the medical sector.
Basically, it's about deciding whether a radiological image, such as an X-ray, shows characteristics of a certain disease. For example, this could be pneumonia, a tumor, or covid19. Based on a learned classification network, the radiologist gets back a value that indicates a possible disease. The problem, however, is that usually the physician does not get any feedback on how this evaluation of the image was generated. But it is precisely this traceability that is particularly important for medicine. Existing approaches visualize areas of images that are used for decision-making. However, these methods are either too imprecise to provide meaningful feedback or ignore the important medical context of the images, which calls into question their trustworthiness. Another important issue is speed. Especially in large clinics or screenings, many images need to be assessed, so time is a relevant factor. Information that is important for a decision has to be delivered in real time in the best case. Not all existing approaches meet these high performance requirements.
For the VRVis method, the researchers involved have extended an existing classification network with another network that learns the features that are important for the decision. This learning process is controlled by "intelligent" and medically correct changes in the image. At the same time, however, this variation in the images also changes the decision about the image. For example, an X-ray image showing a diseased lung automatically becomes an X-ray image showing a healthy lung. This allows the second network to learn which features and areas were important to the decision and then visualize them much more accurately than previously available methods. The VRVis solution scores high in the area of speed and can therefore also be usefully applied in everyday clinical practice. VRVis has already filed two patents on this Interpretable AI solution.