Magnetic Resonance Imaging (MRI) is considered the most versatile and flexible imaging technique in medical diagnostics. Using radiofrequency pulses and magnetic fields, it enables the creation of detailed cross-sectional images of the human body, allowing for a thorough assessment of organs and pathological organ changes. The quality of these images depends largely on the selected setting, referred to as the MRI sequence. This sequence comprises specific hardware commands such as radiofrequency pulses, magnetic field switches, or readout events and is configured according to the requirements of the respective diagnosis. However, configuring these sequences is extremely complicated and requires elaborate, manual programming of hardware commands in complex code. Artificial Intelligence can help identify the optimal sequence for optimal imaging.
To enable this, researchers from the DFKI Research Department Cyber-Physical Systems, the University of Bremen, and Fraunhofer MEVIS are focusing on four aspects in the KIMBi project: developing a machine learning-based MRI simulation for high-quality testing without direct MRI access, reconstructing MRI images from this simulation data while considering novel sequences, optimizing sequence parameters based on expert knowledge for specific examination goals and individual patient characteristics, and developing an application-oriented and easily understandable language for non-programming experts to initiate and adjust the developed optimization.
Artificial intelligence methods can support MRI scans in various aspects. For example, when determining the setting for an MRI scan. Experts call this setting the sequence and protocol. The protocol is the fine-tuning of the sequence, so to speak, in order to precisely meet the needs and characteristics of the respective patient. Or when reconstructing the scan images from the signals received.
Challenges in MRI scans can be categorized into two levels. Firstly, the fit to the situation: Which patient? Which part of the body? What question? And secondly, the technical challenges involved in calculating the image data from the recorded data. Considering that every patient is unique, be it physically, their physique, fat or water content, but also psychologically: Can I lie still for a long time? Does lying still in the MRI stress me out? Am I afraid of the noises in the tube? Additionally, the imaging requirements can vary significantly: Which types of tissue do I want to be able to distinguish best? Which ones interest me less? This highlights the challenge of selecting the most suitable sequence for the situation and the specific patient from numerous options, i.e., the settings for the MRI scan. This requires a great deal of experience and time on the part of the medical staff. Simultaneously, constrained time availability in routine clinical practice, commonly utilized sequences are employed, which often represent a good solution, but could perhaps be improved in order to enable a more comfortable and precise examination of the patient.
We see the potential of the application in several scenarios. On the one hand, there is the classic clinical routine. While radiologists can choose from existing sequences on site to best suit the diagnostic question, it is possible that the optimal sequence is not available on the scanner or does not even exist yet because it does not fit into the usual categories. In these situations, the methodology developed in KIMBi can help to select a precisely fitting sequence in order to improve the overall process. Together with the experience-driven knowledge of medical professionals, it can help to find a better and optimal sequence. At the same time, one important aspect is that teaching is more accessible. Not all complex imaging can be shown and taught everywhere. KIMBI can support this by using AI-controlled simulation and reconstruction to allow sequences to be examined for their properties even without actual scans. Last but not least, the sequence optimization automatically performed by KIMI can also enable the important diagnostics of an MRI in regions with little available specialist personnel.
In our project, the use of AI poses no risk whatsoever. For one thing, physical limits are built into our system so that the algorithm does not make any suggestions outside these limits. More importantly, however, our system's suggestions are always checked by a radiologist before they are carried out on the device. Our system therefore works with the so-called human-in-the-loop. Individuals continue to make decisions and can also contribute their professional expertise. The system provides suggestions, which can be further evaluated by individuals.
This project is a highly interdisciplinary project. Our project partners, the Fraunhofer Mevis and the physics department at the University of Bremen, not only provide the urgently needed domain knowledge for the precise development of our methods, but also develop new methods for the simulation and reconstruction of MRI sequences. At the same time, Fraunhofer Mevis offers the unique opportunity to test MRI sequences in real life. In addition, with the in-house development GammaStar, it offers the possibility of developing manufacturer-independent sequences. DFKI CPS, headed by Rolf Drechsler, is responsible for the development of methods in the field of artificial intelligence and works closely with the cooperation partners.
If you look at the pure concept, a very clear YES. In many areas of application, there is a situation where you want to achieve a specific goal and are looking for the right setting to do so. One example of this is the optimization of alloying and heating processes in steel development. Steels with very specific properties are often sought, for which alloys must then be developed to fit perfectly. However, it is important for every application that the exact characteristics and peculiarities of the respective domain and the specific case must always be taken into account. No two optimization solutions are the same, even if they follow a similar concept. Hence, every new application project presents an exciting challenge.
The KIMBi project is funded as part of the AI Center for Health Care of the U Bremen Research Alliance (UBRA).