How a machine came to a deep learning-based result from a radiology image classified into categories such as “diseased” or “healthy” is not initially comprehensible to the user. Neuronal networks behave like a black box - the learned features of an image that lead to a decision in the network cannot be interpreted by humans.
VRVis has developed methods that make the decision-making basis of a neuronal network visible to the user. These methods not only give medical professionals additional feedback on the reliability of the decision but also help developers to verify the quality of the trained network.
D. Major, D. Lenis, M. Wimmer, G. Sluiter, A. Berg, and K. Bühler, "Interpreting medical image classifiers by optimization based counterfactual impact analysis," IEEE International Symposium on Biomedical Imaging, Iowa City, 2020.
D. Lenis, D. Major, M. Wimmer, A. Berg, G. Sluiter, K. Bühler (2020): "Domain aware medical image classifier interpretation by counterfactual impact analysis," International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2020.
Breast cancer is the most common type of cancer in women worldwide. Thus, the exact analysis of mammography images is an important task in radiology, since the early detection of risk factors and possible suspicious regions in the breast is crucial for successful therapy.
In cooperation with our long-term partner Agfa HealthCare, we developed various deep learning approaches for the analysis of mammography data. We combine several deep learning models that are focused to specific medical questions, such as where pathological lesions and microcalcifications are located in an image or whether a patient has denser breast tissue. This allows us to improve patient-level predictions based on the specific results of the individual models. The modular design of our solution and the possibility to access task specific decisions individually is increasing the explainability of the model and trust in its overall decision. To further increase trust in individual black box deep learning models and their decisions, we developed several novel approaches for explaining AI decisions in the context of image-based decision making.
The exact segmentation of soft tissue tumors in multimodal data is of high importance in various medical domains, e.g. for surgery and biopsy planning or in radiotherapy. For this purpose multimodal data like MRI and PET/CT data is acquired, whereby the tumor segmentation may look different, depending on the modality and clinical task
In cooperation with the Medical University of Vienna, we developed and compared different methods to segment soft tissue tumors. We focused specifically on the modality- and task-specific soft tissue tumor segmentation, and how we can fuse multimodal data efficiently to improve the quality of the segmentation.
T. Neubauer, M. Wimmer, A. Berg, D. Major, D. Lenis, T. Beyer, J. Saponjski, K. Bühler, "Soft Tissue Sarcoma Co-segmentation in Combined MRI and PET/CT Data", Multimodal Learning for Clinical Decision Support Workshop at MICCAI, 2020
In order to support medical professionals in patient screening to detect tuberculosis infections, VRVis has developed a deep learning-based application in collaboration with Agfa Healthcare. With large numbers of patients, it assists in identifying diseased individuals more rapidly, optimising further diagnosis and treatment courses.
According to the World Health Organization, arteriosclerosis ist the no.1 cause of death globally. A detailed analysis of blood vessels in 3D angiograms is generally very time-consuming. In collaboration with Agfa Healthcare, VRVis has developed several methods for efficient semi-automated and fully automated segmentation of peripheral and coronary blood vessels, as well as their semantic annotation; some of these methods have won awards, been published in high-impact journals, patented and integrated into our customers’ products.
A. Novikov, D. Major, M. Wimmer, G. Sluiter, and K. Bühler, “Automated Anatomy-Based Tracking of Systemic Arteries in Arbitrary Field-of-View CTA Scans.,” IEEE Trans. Med. Imaging, vol. 36, no. 6, pp. 1359–1371, 2017.
D. Major, A. Novikov, M. Wimmer, J. Hladuvka, and K. Bühler, “Automated Slice-Based Artery Identification in Various Field-of-View CTA Scans,” presented at the Eurographics Workshop on Visual Computing for Biology and Medicine, 2015, pp. 123–129.
A. Novikov, M. Wimmer, D. Major, and K. Bühler, “A Two-Level Cascade Classification Algorithm for Real-Time Bifurcation Detection in CTA Images of Blood Vessels.,” presented at the Eurographics Workshop on Visual Computing for Biology and Medicine, 2015.
S. Zambal, J. Hladuvka, A. Kanitsar, and K. Bühler, “Shape and Appearance Models for Automatic Coronary Artery Tracking,” In MICCAI 2008 Contest: 3D Segmentation in Clinic: A Grand Challange, 2008. (Gewinner in 2 Kategorien)
VRVis has developed and patented a number of basic methods for medical image analysis. These include widely cited AI methods for simultaneous segmentation of several objects on radiology images, sequential segmentation methods for CT image data, efficient feature detection methods, as well as several basic methods for statistical shape and texture models which, for example, are still used in combination with AI methods for difficult tasks in organ segmentation.
A. Novikov, D. Lenis, D. Major, J. Hladuvka, M. Wimmer, and K. Bühler, “Fully Convolutional Architectures for Multiclass Segmentation in Chest Radiographs,” IEEE Trans. Med. Imaging, vol. 37, no. 8, pp. 1865–1876, Aug. 2018.
J. Beyer, C. Langer, L. Fritz, S. Wolfsberger, and K. Bühler, “Interactive Diffusion Based Smoothing and Segmentation of Volumetric Datasets on Graphics Hardware,” Journal Methods of Information in Medicine 2007, pp. pp. 270–274, 2007.
The spine is an important reference system in the human body for describing the site of pathologies during a radiological examination. For this reason, the anatomical name of each visible vertebra in the image must be known and, in many countries, has to be labelled on the image by radiologists manually. This time-consuming work has to be performed by radiologists in addition to the actual diagnosis within a short space of time. VRVis has developed a group of algorithms which can fully automatically recognise and correctly annotate the anatomy of the human vertebrae from different MR and CT scans independent of the image section. This important automated preparatory work simplifies the radiologist’s work and provides more time for the actual diagnosis.
D. Major, J. Hladuvka, F. Schulze, and K. Bühler, “Automated Landmarking and Labeling of Fully and Partially Scanned Spinal Columns in CT Images,” Med Image Anal, vol. 17, no. 8, pp. pp. 1151–1163, 2013.
M. Wimmer, D. Major, A. A. Novikov, and K. Bühler, “Fully automatic cross-modality localization and labeling of vertebral bodies and intervertebral discs in 3D spinal images.,” Int. J. Computer Assisted Radiology and Surgery, vol. 13, no. 10, pp. 1591–1603, 2018.
J. Hladuvka, D. Major, and K. Bühler, “Bone Profiles: Simple, Fast, and Reliable Spine Localization in CT Scans,” in Lecture Notes in Computational Vision and Biomechanics, vol. 20, no. 15, Cham: Springer International Publishing, 2015, pp. 173–184.
Today, cardiovascular diseases are the most common cause of death in Austria. A detailed analysis of (time-dependent) 3D image data of the beating heart and coronary arteries in a patient is extremely time-consuming. Fully automated methods which simplify the quantification of heart pathologies can therefore make a considerable contribution to accelerating the diagnosis by the radiologist. VRVis has already developed several solutions for segmenting the heart ventricles on CT and MR images as well as an award-winning solution for fully automated segmentation of coronary arteries.
S. Zambal, A. Schöllhuber, K. Bühler, and J. Hladuvka, “Fast and robust localization of the heart in cardiac MRI series,” in Proceedings of the International Conference on Computer Vision Theory and Applications (Visapp 2008), pp. pp. 314–316, 2008.
S. Zambal, J. Hladuvka, and K. Bühler, “A fully automatic system for segmentation and analysis of the left and right ventricles of the heart using a bi-temporal two-component model,” in Proceedings of CARS 2007 (Computer Assisted Radiology Surgery), 2007.
In association with our application-oriented research work for our business partner Agfa Healthcare, many patents for AI-based solutions have been produced in our in-house basic research, for example on the fully automated segmentation and annotation of blood vessels or human vertebrae. You can find out more about our patents.
The long-term vision of this applied research project is to use available data resources to improve image-based diagnostics based on complex data in daily clinical routine.
The strategic project forms the organizational and scientific hub for the realization of an area wide integrative visual computing approach. It covers joint strategic research and development on fundamental challenges in all application projects.
VRVis a founding member of the Austrian BioImaging/CMI, which is a professional consortium of multiple Austrian science institutions and the official Austrian Euro-BioImaging initiative.
COMULIS is an EU-funded COST Action that aims at fueling collaborations in the field of correlated multimodal imaging (CMI).
Visual computing techniques for the automated detection of osteoporosis and osteoarthritis.
Visual computing for medicine: image processing solutions for new applications in radiology.
The analysis, visualization and exploration of high-dimensional image spaces are the subject of the KAFus project.
Next generation workflows for interactive knowledge generation from images and simulations.
Software for the use of multi-modality images in external radiotherapy.
On March 3, 2020, Katja Bühler, head of our Biomedical Image Informatics Group, was awarded with the renowned TU Women's Prize.
Visual computing for computer-aided diagnostics and operation planning.
Sebastian Zambal, Jiří Hladůvka, Armin Kanitsar, Katja Bühler, Shape and Appearance Models for Automatic Coronary Artery Tracking, WON MICCAI 2008 Contest: 3D Segmentation in Clinic: A Grand Challange.
C. Langer, M. Hadwiger, K. Bühler, Interaktive diffusionsbasierte Segmentierung von Volumendaten auf Grafikhardware, Bildverarbeitung für die Medizin 2005; GI Informatik Aktuell; Springer Verlag. pp 168-17, BVM 2005 Best Poster