How do I find an algorithm expert

How Artificial Intelligence analyzes microscopic images

Modern microscopy methods provide huge amounts of image information. Whether genes or nervous systems are being examined - the precise technology is a blessing for medical research and diagnostics. However, a well-founded expert analysis is required to evaluate such recordings. A complex, time-consuming, but necessary task for scientists.

A research team from the Julius Maximilians University of Würzburg (JMU) and the University Hospital of Würzburg (UKW) now want to break new ground: They not only want the images to be analyzed automatically by artificial intelligence (AI), but also the quality with the help of AI improve the image analysis. To this end, they wrote a study and developed new guidelines on how machine learning can make expert-based image analysis more objective and valid.

"In our work we asked ourselves how microscopic images can be automated and objectively analyzed with the help of so-called deep learning algorithms from AI," explains Robert Blum, neurobiologist at the Institute for Clinical Neurobiology at the FM. For this purpose, an interdisciplinary team of neuroscientists from the UKW and business informatics from the JMU has been formed to examine this question in practice.

More experts, more objectivity

The result: if the self-learning AI algorithms train with the data from a single expert, this can lead to the AI ​​learning the subjective analysis criteria of the expert. Objectivity can fall by the wayside. “However, if you use the common knowledge of many experts to train an algorithm, it is less susceptible to subjective analysis criteria. This makes the evaluation of image data more objective and reproducible, ”says Blum.

This was checked by comparing the image analyzes of several experts. In order to clarify experimentally how valid artificial neural networks can work, neurobiological laboratory experiments produced objective comparison parameters. The results have recently been published in the specialist journal “eLife”.

The team makes it clear that the assessment of image data by experts is still the "gold standard in research and clinical practice". "But we have shown that by embedding artificial neural networks in a structured workflow, the analysis of image data can not only be automated, but is also objectively and reliably possible," explains Christoph Flath, Chair of Information Systems and Information Management at JMU.

Use also in less developed countries

Before these AI algorithms can also take over the analysis and evaluation of microscopic images for science and clinics, it will take time - and more research.

"In order to make access to our deep learning analysis process as barrier-free as possible, we are currently working on the development of a user-friendly toolbox that is based on the guidelines we have developed for objective image analysis," says Flath. The goal: to make the data and trained algorithms available to every researcher or interested party with internet access. "We see particular opportunities here for less developed countries," explains Blum.

Initiative from doctoral students

The research idea goes back to the doctoral students Matthias Griebel (Business Informatics, JMU) and Dennis Segebarth (Clinical Neurobiology, UKW), who designed and carried out the project together with Blum and Flath.

The team was funded by the German Research Foundation as part of a special research area TRR58 for 'Fear, Anxiety, and Anxiety Diseases', the Graduate School of Life Sciences and the Interdisciplinary Center for Clinical Research at JMU, as well as the Austrian Fund for the Promotion of Scientific Research. The work was also supported by researchers from the Universities of Münster and Innsbruck.

publication

Segebarth et al .: "On the objectivity, reliability, and validity of deep learning enabled bioimage analyzes", in: eLife, DOI: 10.7554 / eLife.59780

Contact

Prof. Dr. Christoph Flath, Chair for Information Systems and Information Management, University of Würzburg, T +49 931 - 31 85128, [email protected]

PD Dr. Robert Blum, Institute for Clinical Neurobiology, University Hospital Würzburg, T +49 931 - 201 44031, [email protected]

By Kristian Lozina

Back