Medical imaging generates a lot of data, for example during computer tomography. This data important when it comes to personalised medicine. Artificial intelligence methods, such as machine learning, use this data to learn and help tailor diagnoses and therapies to individual needs in the future. However, such systems are currently still burdened with uncertainties. A team of researchers from Kaiserslautern and Leipzig is working on a system that automatically analyses and visualises medical data, including their uncertainties. The researchers will be presenting this technology at the medical technology trade fair “Medica” in Düsseldorf held from 14 to 17 November at the Rhineland-Palatinate research stand (stand E80, hall 3).
In the event of a stroke, speed is of essence. Using computer tomography (CT) scans, doctors can quickly determine the position of the blood clot in the brain and what treatment is appropriate. Such imaging procedures play an important role in medicine. They are also used in other areas, for example prior to operations. Magnetic resonance imaging (MRI) scans help surgeons plan an operation before it is performed.
What all these techniques have in common is that they generate a lot of data. “Analysing and visualising this data automatically is an important step toward personalised medicine,” says Dr. Christina Gillmann, a computer scientist at the University of Leipzig. “This area has gained a lot of importance in recent years.” AI processes such as machine learning and neural networks make this possible. These networks learn on the basis of data with which they are trained or “fed”. They learn, for example, from CT image data a doctor has previously processed. In this way, technical information, but also medical experience, is incorporated. The rule is that the more data these methods can evaluate, the better the results will be.
In a few years, such technologies have the potential to be used in everyday clinical practice, for example, to enable personalised diagnoses and therapies. However, they are still in the early stages of development. “Each medical case has to be trained individually. The data must be prepared individually in advance, which is very time-consuming,” explains Robin Maack from the Computer Graphics and Human Computer Interaction working group at Technische Universität Kaiserslautern as a problem. For example, doctors have to “label” the data for each medical case individually. “This means, for example, that if a network is to train to automatically recognise a tumour, hundreds of images with known tumours have to be hand-drawn in so that the neural network has a basis with which to learn,” Gillmann explains.
Maack continues: “In addition, there are no standard interfaces with which trained networks can be handled, loaded and used. But also when there are uncertainties in the data layers; be it training data sets or models used; there are no standardised guidelines for how medical professionals should deal with that.”
Such uncertainties occur, for example, with lesions. During a stroke, certain areas of the brain are no longer supplied with sufficient oxygen, or not at all, due to the blockage of vessels in the brain. They are no longer able to work efficiently. The core of the lesion is often easy to recognise, but at the edge there is usually no clear demarcation and regions where even doctors cannot agree whether they should be classified as a lesion or not. Ultimately, what is needed here is medical experience on how to deal with these issues.
This is where the focus of Gillmann and Maack's work starts. Their team is currently developing a uniform system for processing and evaluating medical image data and visualising their uncertainties. The system is called GUARDIAN. The researchers have designed their technology in such a way that it is easy to use. “Clinics can load their trained neural networks and combine them with the processed data provided, for example, in the case of a stroke”.
The system evaluates the data and visualises the results. “This happens automatically, without the need for IT knowledge,” Maack continues. “In addition, our technology also shows the uncertainties.” This means that the doctors can look at them again and, if necessary, make a joint decision on what is the best treatment in an individual case, for example.
The system is freely available as an open-source application. The two computer scientists will be presenting this system at the fair.
The “Computer Graphics and Human Computer Interaction” working group of Professor Dr. Hans Hagen at TU Kaiserslautern has been researching for a long time to prepare data from imaging procedures for medicine in such a way that it can be used simply and reliably in everyday clinical practice.
Klaus Dosch, Department of Technology and Innovation, is organizing the presentation of the researchers of the TU Kaiserslautern at the Medica. He is the contact partner for companies and, among other things, establishes contacts to science.
Contact: Klaus Dosch, E-mail: firstname.lastname@example.org, Phone: +49 631 205-3001
Questions can be directed to:
Computer Graphics and Human Computer Interaction
Phone: +49 631 205-3268
Image and Signal Processing Group
Universality of Leipzig
Tel: +49 341 97 32281