AI for stroke diagnostics

The joint project KI-SIGS is dedicated to the development of an "AI Space for Intelligent Health Systems". As part of the project, mbits is working with the University Hospital Schleswig-Holstein (UKSH) to develop AI-based solutions that support radiologists in the diagnosis of computer tomography (CT) images of stroke patients.

In the case of a stroke, every minute counts, but at the same time the quality of care must not suffer under time pressure. This requirement affects the entire stroke rescue chain, from initial care and diagnosis to the choice of optimal therapy. Within 15min of the patient's arrival, a CT must be run to make the diagnosis. This can rule out cerebral hemorrhage, localize occlusions in the cerebral arteries, and quantify the resulting ischemia. In the future, AI algorithms will assist radiologists in making findings. Together with physicists and stroke experts from radiology and neuroradiology at UKSH, mbits is researching such algorithms. In the project, two questions 1) exclusion of cerebral hemorrhages and 2) localization of occlusions are addressed. Solutions based on native CT are to be developed for both tasks, which offers the advantage that no contrast medium is required for image acquisition. This article will address the first question in more detail, a follow-up article will address the second question.

Figure 1: Cerebral hemorrhage in native CT.

Despite similar symptoms, the therapies for hemorrhage and stroke are fundamentally different and even contraindicate each other. If bleeding occurs, it must be stopped and, depending on the extent, treated neurosurgically. On the other hand, in case of stroke, the occluded vessel must be reopened to restore blood flow to the infarcted brain tissue. In many cases, this is achieved by thrombolysis, in which anticoagulants are given to dissolve the thrombus. However, such medication would have serious consequences for the patient in the case of an acute cerebral hemorrhage. Accordingly, hemorrhage exclusion in native CT is a critical step, which we aim to support by an AI algorithm.

Figure 2:
Four-layer convolutional neural network for classification of a 3D CT image.

Deep learning methods are used for this purpose. Deep learning is a form of AI in which neural networks are used to process data. For image processing, In recent years, convolutional neural networks (CNNs) have become established for image processing. These consist of several layers in which the input image is filtered. By stringing together filters, the network can learn both simple features, such as edges, and complex patterns, such as faces or the appearance of brain hemorrhages. Based on these learned features, the network can then assign whether or not there is bleeding in the present CT.

Alexandra Ertl | Research & Development

Newsletter

Would you like to receive regular information about new products, mRay updates and clear tutorial videos on new features? Subscribe to our quarterly newsletter.

This might also be interesting for you

EN

Vielen Dank für Ihre Anmeldung

In Kürze erhalten Sie eine E-Mail, mit der Sie das Abonnement unseres Newsletters bestätigen können.