Data for the implementation of innovative AI products is often not big enough though and also not representative of the general public. In addition, medical image data are highly sensitive patient data which are subject to the strict regulations of the General Data Protection Regulation (GDPR) and cannot be used without restriction.
This is exactly where the new “NeuroTest” project at Landshut University of Applied Sciences, overseen by Prof. Stefanie Remmele, is stepping in. In collaboration with the Munich-based medical technology company deepc, the professor of Medical Technology research is researching how artificial patient data can be developed for use in AI models in medical imaging. At the same time, the project partners are working on an online platform to offer manufacturers of medical devices the opportunity to test their AI-based medical devices before applying for a license. The project is being funded by the Federal Ministry for Economic Affairs and Energy with more than 400,000 euros.
In the next two years, the project team at Landshut University of Applied Sciences will be looking at the conditions under which AI models can deliver a consistent and meaningful result in imaging, in order to support doctors in the diagnosis accordingly. Unlike conventional methods, the accuracy of the AI solution not only depends on the data processing logic, but also on the data used to train the technology. "This is particularly challenging when processing MRI data, because contrast and image quality can vary greatly, there is no infinite number of images and the available training images never cover the entire range of possible variations," Remmele explains.
Landshut University of Applied Sciences is researching the influencing parameters of human brain images using existing MRI images and artificial test images. The patient's age, gender, pre-existing illnesses or genetic and ethical information play an important role here, as do fluctuations in the recording parameters and the I hardware. "With the help of this generated and analyzed data and the model knowledge about technical influences, we want to create artificial data sets which variations can be simulated from depending on the hardware, findings or patient," Remmele explains the procedure. "With this, we can then test AI models against all these variations and detect which changes in the data a model does not react sufficiently robustly to," says Remmele. The major challenge is to standardize the generated data in such a way that AI models can be evaluated sustainably and do not provide any incorrect information.
At the same time, deepc's AI and software specialists are developing software that enables manufacturers, for example PAC systems, to validate their AI-based products online and achieve the specified safety standards. "With the help of the development of methods to create synthetic reference data in combination with real patient data, which are constantly checked, added to and compared at the same time using a standardized software platform, we expect clear progress in the standardization, application and especially in the approval process for AI solutions in the field of imaging medical technology," explains Dr. Franz Pfister, CEO of deepc.
MEDICA-tradefair.com; Source: Landshut University of Applied Sciences