Medical imaging through magnetic resonance imaging (MRI) typically takes a lot of time since each image must be created from a collection of individual measurements. However, with the aid of machine learning, it is now possible to generate images using fewer MRI measurements, which can reduce both time and costs. The crucial requirement for this approach is having flawless images for training AI models, which are often not available for specific uses, such as real-time MRI, where images tend to be somewhat blurred. Recently, an international research group has achieved the creation of accurate live MRI images of a beating heart without needing perfect training images and using minimal MRI data by leveraging cleverly trained neural networks. This advancement may lead to more frequent practical applications of real-time MRI in the future.
Thanks to the innovative training of neural networks, researchers at TU Graz have accomplished the generation of accurate real-time images of the beating heart using only a small amount of MRI measurement data. This new technique also has the potential to enhance other MRI applications.
Medical imaging through magnetic resonance imaging (MRI) is typically quite slow because each image requires combining data from multiple individual measurements. The introduction of machine learning has enabled imaging with fewer MRI data points, helping to save both time and costs. However, a key condition for this process is the availability of impeccable images for AI model training. Unfortunately, such ideal training images are absent in specific contexts, like real-time MRI, where the images invariably display some blurriness. An international research team headed by Martin Uecker and Moritz Blumenthal at the Institute of Biomedical Imaging, Graz University of Technology (TU Graz), has recently succeeded in producing accurate live MRI images of the heart, even without perfect training images and with limited MRI data through the use of intelligently trained neural networks. As a result of these advancements, we may see a rise in the practical use of real-time MRI in the years to come.
Optimizing imaging through withheld data
To train their machine learning model for MRI imaging, Martin Uecker and Moritz Blumenthal employed self-supervised learning techniques. Rather than relying on pre-curated perfect images, they based their training method on a portion of the original measurement data intended for image reconstruction. Moritz Blumenthal elaborates, stating, “We divided the MRI data into two parts. Our machine learning model reconstructs the image using the first, larger portion. It then attempts to predict the second portion of the data that was withheld from it based on the reconstructed image.” If the model struggles or performs poorly in this prediction, it indicates that the previously reconstructed image was likely incorrect. Consequently, the model is refined, producing a new improved image version and retrying to estimate the second data portion. This iterative process continues until the results become reliable. Through this training routine, the system learns the characteristics of high-quality MRI images, enabling the model to directly produce a quality image in practical applications.
This method can accelerate and reduce costs for various MRI applications
“Our method is ready for implementation,” states Martin Uecker, “although it might take some time before it shows up in clinical practice.” This technique is applicable to many other MRI uses, streamlining them and making them more cost-effective. For instance, in quantitative MRI, where precise measurements and quantifications of tissue parameters are necessary, radiologists can access accurate data for diagnoses rather than relying on visual assessments based on brightness variations. Martin Uecker notes, “Previously, quantitative MRI procedures often required substantial time. With our machine learning model, we’ve succeeded in significantly speeding up these measurements without sacrificing quality.”
The findings from this research, published in the journal Magnetic Resonance in Medicine, stem from a collaborative international and interdisciplinary effort at the Institute of Biomedical Imaging. Contributors included Christina Unterberg (a cardiologist at the University Medical Centre Göttingen), Markus Haltmeier (a mathematician from the University of Innsbruck), Xiaoqing Wang (an MRI researcher from Harvard Medical School), and Chiara Fantinato (an Erasmus student from Italy). The algorithms and MRI data are made publicly accessible, allowing other researchers to replicate the results and build upon this new methodology.
