Artificial intelligence is becoming increasingly important in medicine. Researchers in Germany are pursuing various projects to explore how algorithms might assist medical personnel.
Doctors have to analyse images all the time in their daily work. In the future, artificial intelligence (AI) methods will help them do so more quickly, more reliably and in greater detail. Researchers at the German Cancer Research Center (DKFZ) in Heidelberg have developed a new method that allows AI methods to be applied automatically to many different data sets. This will also improve the outlook for future application in clinical settings.
Named nnU-Net, the new program was developed as part of the doctoral thesis written by medical IT expert Dr Fabian Isensee, who is a member of Professor Klaus H. Maier-Hein’s team at the DKFZ. The paper in which the program was presented appeared in the renowned journal Nature Methods and was written by Isensee with his colleague, the AI specialist Dr Paul Jäger.
The program is used to “segment” 3D imaging data. This involves the algorithm scanning an image to find all the pixels that relate to a particular object. A good example of imaging data that need to be segmented are computed tomography images. To be able to interpret them, clinicians first have to identify the organs and distinguish them from other objects, such as tumours, for instance.
Generally speaking, the AI software is taught how to segment on the basis of example data. In medicine, imaging data come from all kinds of instruments and differ considerably in terms of their properties, e.g. their size and resolution. The algorithm therefore has to be repeatedly optimised and adapted to specific data sets before the AI software can deliver meaningful results. The latest nnU-Net program now handles most settings completely automatically.
Using algorithms to segment the data saves doctors time, explains Isensee. One example is radiation therapy: automatic methods allow the segmentation to be carried out quickly for all organs, meaning that more precise planning is possible. This also allows key organs to be exposed to lower levels of radiation.
To date, nnU-Net has been used solely in research settings. It will be a long time before it is approved for clinical practice. Nonetheless, Isensee believes it is realistic to assume that radiologists and pathologists will get help from AI in the not too distant future.
AI methods can help doctors in other areas, too. In Northern Germany, for example, a joint research project called “KI Space for Intelligent Health Systems” is currently being launched. AI institutes in Bremen, Hamburg and Schleswig-Holstein are involved, alongside medical technology companies and partners of university hospitals. The objective is to develop adaptive medical systems, learning robot assistants and further similar aids for healthcare. In a cooperation project with the Austrian Software Competence Center Hagenberg, the scientists are pursuing an information theory approach to developing trustworthy AI.
In addition, researchers at the “International Future Laboratory for Artificial Intelligence in Hannover” want to develop new solutions for personalised medicine with the aid of AI. To this end, they are collaborating with researchers in Australia, New Zealand, Singapore, India, the USA and Europe. According to the website of the future laboratory: “Scientific excellence thrives on exchange with the best in the world”.
Applying AI in medicine can entail great advantages for doctors and patients alike. And it is by no means limited to imaging data – even if considerable success has already been achieved in this area.