10-fold Deep Learning Accelerated MRI of Joints
Final Report Abstract
Magnetic resonance imaging (MRI) has become an essential tool for the non-invasive diagnosis of joint diseases, offering high-resolution visualization of both acute and chronic pathologies. Two-dimensional (2D) turbo spin-echo (TSE) sequences are considered the clinical standard for imaging intra-articular structures due to their high signal-to-noise ratio and excellent in-plane and contrast resolution. However, a major limitation of conventional MRI is the relatively long acquisition time. Accelerating image acquisition is therefore highly desirable in clinical routine - to improve patient comfort, reduce motion artifacts, and increase economic efficiency. Several acceleration techniques are already established, including parallel imaging (PI), simultaneous multi-slice (SMS) acquisition, and compressed sensing (CS). The combination of PI and SMS has proven particularly effective, as their effects are multiplicative, allowing for higher acceleration factors. A clinical implementation of a fourfold accelerated 2D TSE MRI of the knee (SMSx2-PIx2) has already been demonstrated. However, further acceleration is limited by factors such as specific absorption rate (SAR), artifacts, and the inherent constraints of conventional reconstruction methods when dealing with signal-starved raw data. Deep learning (DL)-based reconstruction methods show great potential for further accelerating MRI. These approaches enable superior noise suppression and artifact reduction, allowing for substantially faster scans while maintaining - or even improving - image quality. For example, a DL-based super-resolution technique applied to fourfold accelerated PI-SMS knee MRI has demonstrated improved detection of cartilage lesions compared to conventional MRI. As part of the fellowship project, existing deep learning (DL)-based reconstruction methods were systematically evaluated to assess their performance and limitations regarding higher acceleration factors in MRI. In addition, new DL-based approaches were developed and specifically optimized to enable eight-fold or higher acceleration - an objective that could not be achieved artifact-free with previously available DL techniques. Furthermore, in three additional subprojects, DL-based methods were successfully applied to accelerate 3D MRI sequences. This ultimately enabled up to nine-fold accelerated ankle MRI, with a total acquisition time of under two minutes for a full 3D TSE sequence.
Publications
-
8-fold Accelerated TSE MRI: First Clinical Assessment of a Dedicated End-to-End DL Image Reconstruction Approach for Combined Simultaneous Multislice and Parallel Imaging Acceleration. Annual Meeting Abstracts of the German Society of Skeletal Radiology (DGMSR) 2025, May 16–17, Berlin/Germany. Skeletal Radiology, 54(S1), 1-6.
Leonhardt Y., Vosshenrich J., Mostapha M., Koerzdoerfer G., Raithel E., Nadar M., Bruno M. & Fritz J.
-
8-fold Accelerated TSE MRI: First Clinical Results of an End-to-End DL Reconstruction Method for Combined SMS and Parallel Imaging Acceleration. ISMRM Annual Meeting. ISMRM.
Leonhardt, Yannik; Vosshenrich, Jan; Mostapha, Mahmoud; Koerzdoerfer, Gregor; Raithel, Esther; Nadar, Mariappan; Bruno, Mary & Fritz, Jan
-
Arthroscopy-validated Diagnostic Performance of 7-Minute Five-Sequence Deep Learning Super-Resolution 3-T Shoulder MRI. Radiology, 314(2).
Vosshenrich, Jan; Bruno, Mary; Cantarelli Rodrigues, Tatiane; Donners, Ricardo; Jardon, Meghan; Leonhardt, Yannik; Neumann, Shana G.; Recht, Michael; Serfaty, Aline; Stern, Steven E. & Fritz, Jan
-
MR Neurography of the Foot and Ankle Nerves. Magnetic Resonance Imaging Clinics of North America, 33(3), 545-561.
Leonhardt, Yannik & Fritz, Jan
