Artificial intelligence can accelerate the detection of pathological tissue in medical image processing.
Deepfake is a way of stalking or rushing or of making a new film life for long-dead screen stars like James Dean. One can smile at these images or videos manipulated with artificial intelligence and at the same time fear them and their consequences for society and politics. Deepfake can swap faces in videos and put words in the mouths of politicians they never said, or incorporate pop stars into scenes. Deepfake thus poses a great danger to the credibility of moving images, which is why the discussion about this is mainly filled with negative omens. Not so in the course of study Information Technology & Systems Management of the University of Applied Sciences Salzburg. “We are using a similar technique for medical image processing with the aim of improving diagnoses in medicine,” explains Michael Gadermayr, senior lecturer in this field, in an interview with the “Salzburger Nachrichten”. The aim of this research is to solve medical questions by means of artificial intelligence, for example, to make deepfake technologies usable for the detection of sick body tissue such as tumors. Specifically, Gadermayr’s research focuses on the automated processing of image data in digital pathology. Artificial intelligence has the potential to relieve doctors of visually demanding diagnoses in the future: “There is a huge amount of data for diagnostics, for example cancers. The microscopic images, for example, have resolutions in the range of up to several gigapixels.”
A thorough manual analysis of this data, including all the information involved, is extremely time-consuming or not feasible – here, data preprocessing by artificial intelligence can provide important support. “With detection (detection), segmentation (decomposition) and measurement of relevant tissue areas, we can perform a time-efficient data-driven analysis,” says the researcher. Since his dissertation in computer science at the University of Salzburg and other research projects, the native of Oberalmer has at RWTH Aachen University. Now he is in charge of setting up this research at the University of Applied Sciences Salzburg. The research project “Artificial Intelligence for the Analysis of Medical Image Data (KiaMed)”, which he is conducting together with ge’o’te o’a-year-old director of biomedical analytics at the FH, was recently approved through state funding. Innovative imaging techniques can revolutionize the analysis of medical image data and thus subsequently revolutionize diagnostics, promote diagnostic accuracy and patient safety. At KiaMed, artificial intelligence replaces manual data entry by physicians, saving time and money. However, the precision of the pixel-accurate analysis of medical image data by artificial intelligence is largely related to the availability of training data. “Our focus is on data from magnetic resonance therapy (MRI), computed tomography (CT) and histology,” says Gadermayr. “The advantage of these modalities is that they are highly standardised. This makes the data from different clinics more comparable, which is then important for diagnostics.” Because a big hurdle is the diversity of the data. Compared to humans, however, the machines cannot compensate for structural differences between training and test data by means of “natural intelligence”. “The computer is relatively stupid in terms of generalization,” the researcher says.
The artificial neural networks therefore need training: “For this purpose, we use the same technology as Deepfake.” A reward or punishment system trains the nets so that they can produce deceptively real-looking images or distinguish them despite their similarity. This creates virtual pathological sectional images that can be processed like original images and improve diagnostic accuracy. They also work with physicians in the region, such as the specialists at the Salzburg State Hospital or the Cardinal Schwarzenberg Hospital in Schwarzach. The medical profession is now very open to medical image processing using artificial intelligence. There is also no danger, says Gadermayr, that “Dr. Deepfake” will take work or expertise away from doctors in the foreseeable future: “Our goal is not to replace the physicians, but to increase the efficiency and accuracy of diagnostics with computer-aided methods. We are still a long way from a time when there is no longer any need for doctors and only one has to go to the machine.”