How should algorithms be investigated

Recognizing diseases with algorithms

Automated image analysis

Many doctors are under time pressure and want more time for the patient. Intelligent computer programs could help here: students of biomedical engineering at the Landshut University of Applied Sciences researched how algorithms can support radiologists in examining medical images as part of a project.

Thanks to X-rays, ultrasound or magnetic resonance, doctors can discover diseases in the body without surgery. To do this, radiologists have to analyze numerous images - this takes a lot of time and concentration. In the future, computers could make their work easier. The Landshut students Jakob Dexl, Lisa-Marie Kirchner, Maximilian Reiser and Michael Uhl of the Biomedical Engineering course examined the extent to which certain algorithms can presort cranial MRI images. That would mean doctors can focus more on suspicious images with difficult findings. The Radiology Mühleninsel practice from Landshut provided real MRI data from patients - completely anonymized, of course. The team also received support from radiologist Prof. Dr. Andreas Lienemann and in the technical area from the health IT company Cerner Deutschland GmbH.

Machine learning: the program learns with you

Research has been going on in the area of ​​machine learning for years. The topic is also emerging in medical technology

Maximilian Reiser

For the automatic classification of images into “sick” and “healthy”, modern machine learning algorithms are used, which, like a radiologist, train the classification using existing data. “A major challenge in this project was that images of healthy patients also differ greatly from one another, for example due to benign diseases or age. So the heterogeneity is huge, ”explains professor Prof. Dr. Stefanie Remmele, who supervised the team in their project and final theses. “At the same time, the radiologist sometimes only differentiates between sick and healthy on the basis of tiny details. That makes it very difficult for the algorithm to classify the images correctly. ”The students therefore defined image details and features that the program should use to distinguish between healthy and sick - for example, the shapes and sizes of brain structures. To do this, they fed the algorithm with images and data.

In the first test as part of the study projects, it did well. But before the computer-aided diagnosis can really relieve radiologists in their work, "a lot of research will be necessary on the subject," says Dexl. For example, he and his team only used one image per patient - in reality, more images are recorded for one finding, such as different sections of the brain or different views. Dexl's colleague Reiser is certain: “Research in the area of ​​machine learning has been going on for years. The topic is also emerging in medical technology. "

Source: Landshut University of Applied Sciences