Heidelberg – Scientists from the Heidelberg University Hospital, the German Cancer Research Center and the Bonn University Hospital have developed an algorithm for the automated evaluation of CT images. The system is clearly superior to several commercially available AI products for stroke diagnosis. For research purposes, the algorithm can be used free of charge via the website.
Sudden, persistent tingling in your arm or an unusually severe headache? The question of whether it is a stroke is sometimes not easy to answer, even after computed tomography (CT). A new analysis program developed by teams from the University Hospital Heidelberg (UKHD) and the German Cancer Research Center (DKFZ) in cooperation with colleagues from the University Hospital Bonn (UKB) offers support in diagnosing – or ruling out – strokes. The learning algorithm was trained with the CT datasets of more than 1,000 UKHD patients with suspected stroke and tested for reliability in three regional clinics and at the UKB in parallel with the usual stroke diagnostics and compared with commercial products. The results of the test runs have now been published in the journal “Nature Communications”: The Heidelberg application for evaluating CT image data performs significantly better than currently available AI products for stroke diagnostics.
Computed tomography has become indispensable in diagnosing a stroke: the examination is quick and provides a view of the causative vascular occlusions, their localization and extent, as well as the tissue damage in the brain. This information is important for the further procedure, eg whether a catheter intervention, the so-called thrombectomy, is an option and the patient has to be transferred to a specialized clinic like the UKHD. The assessment of the CT images has to be quick, because the more time it takes to start therapy, the more brain tissue is irreparably damaged by the circulatory disorder. “In the case of rare occlusions in the outer brain areas, however, the assessment of the image data can take time and is subject to a certain degree of uncertainty, which also depends on the experience of the team on site,” says Philipp Vollmuth, senior author of the article and head of the section Computational Neuroimaging, UKHD Clinic for Neuroradiology.
Computer-aided analysis programs are now on the market as time-saving support. For this purpose, a learning algorithm is trained on the basis of as much CT image data as possible to identify vascular occlusions of any position and type in the brain as reliably as possible. The problem with commercial providers: they only have limited access to clinical data, the artificial intelligence (AI) training material is therefore not always representative and the methodology used does not always correspond to the current state of science. As a result, the AI poorly recognizes occlusions in brain vessels that are less frequently affected.
The Heidelberg analysis tool is different: The algorithm was developed using image data from more than 1,000 patients who were examined and treated at the UKHD with suspected stroke. It then proved itself in a six-month test phase at three selected partner clinics of the Rhine-Neckar Stroke Consortium (FAST) and at the UKB. “Our algorithm proved to be significantly more accurate compared to commercial AI products. In addition, the negative predictive value, i.e. that the prediction is actually correct when the algorithm rules out a vascular occlusion, is already very good at up to 97 percent,” says Philipp Vollmuth. “But our goal is to make the algorithm even better. “
The project team wants to use “data crowdsourcing” for this: The algorithm can be used via the website for research purposes and is made available free of charge by the DKFZ. Registered researchers can upload the image data of patients to the website and receive the evaluation within a few minutes Image data should be used in the future after approval to continue training the AI.”With the help of this concept, we will continuously improve the algorithm and therefore hope that as many clinics as possible will participate in the project. We see this concept as a blueprint for the cooperative Development of highly reliable AI algorithms,” explains Ralf Floca, group leader in the Department of Medical Image Processing at the DKFZ
The AI cannot and should not replace the critical look of a doctor: the four-eyes principle still applies to the diagnosis. “The system helps to confirm the diagnosis, especially when the symptoms are not clear and there is no specialized neuroradiological expertise in smaller hospitals. In addition, our algorithm indicates how reliable it is in its result. If there is great uncertainty, further investigations should follow,” explains Philipp Vollmuth. He sees a great benefit for those affected above all in the possibility of being able to initiate therapy more quickly in this time-critical situation thanks to AI support.
Literature:
Brugnara G, Baumgartner M, Scholze ED, Deike-Hofmann K, Kades K, Scherer J, Denner S, Meredig H, Rastogi A, Mahmutoglu MA, Ulfert C, Neuberger U, Schönenberger S, Schlamp K, Bendella Z, Pinetz T, Schmeel C, Wick W, Ringleb PA, Floca R, Möhlenbruch M, Radbruch A, Bendszus M, Maier-Hein K, Vollmuth P. Deep-learning based detection of vessel occlusions on CT-angiography in patients with suspected acute ischemic stroke.
Nature Communications 2023,
Additional Information:
Source: Press release from Heidelberg University Hospital
With more than 3,000 employees, the German Cancer Research Center (DKFZ) is the largest biomedical research facility in Germany. Scientists at the DKFZ research how cancer develops, record cancer risk factors and search for new strategies to prevent people from developing cancer. They are developing new methods with which tumors can be diagnosed more precisely and cancer patients can be treated more successfully. The Cancer Information Service (KID) of the DKFZ provides those affected, interested parties and specialist groups with individual answers to all questions on the subject of cancer.
In order to transfer promising approaches from cancer research to the clinic and thus improve the chances of patients, the DKFZ operates translation centers together with excellent university clinics and research institutions throughout Germany:
National Center for Tumor Diseases (NCT, 6 locations) German Consortium for Translational Cancer Research (DKTK, 8 locations) Hopp Children’s Cancer Center (KiTZ) Heidelberg Helmholtz Institute for Translational Oncology (HI-TRON) Mainz – a Helmholtz Institute of the DKFZ DKFZ-Hector Cancer Institute at the University Medicine Mannheim National Cancer Prevention Center (together with the German Cancer Aid)
The DKFZ is funded 90 percent by the Federal Ministry of Education and Research and 10 percent by the state of Baden-Württemberg and is a member of the Helmholtz Association of German Research Centers.