Accelerated Deep-Learning-Based Image Reconstruction for 3D T2 Dark-Fluid in Imaging of Multiple Sclerosis.
Authors
Affiliations (1)
Affiliations (1)
- Department of Neuroradiology, University Medical Center Mainz, Johannes Gutenberg University, Langenbeckstr (S.A., M.A.A.M., A.E.M., M.K., V.I.S., M.A.B., A.E.O.); Institute of Medical Biostatistics, Epidemiology and Informatics (IMBEI), University Medical Center Mainz, Johannes Gutenberg University, Rhabanusstr, Mainz (R.H.P.); Research & Clinical Translation, Magnetic Resonance, Siemens Healthineers AG, Erlangen (M.D.N.); Department of Neurology, University Medical Center Mainz, Johannes Gutenberg University, Langenbeckstr, Mainz, Germany (V.F.).
Abstract
Deep-learning (DL)-accelerated MRI can significantly reduce acquisition times. Studies evaluating interchangeability with conventional 3D data sets, particularly for monitoring disease activity in multiple sclerosis (MS), are lacking. This study investigated interchangeability and comparability between conventional fluid-attenuated 3D-T2-SPACE dark-fluid (c-3D-T2) and accelerated DL-based fluid-attenuated 3D-T2-SPACE dark-fluid (DL-3D-T2) in detecting new white matter lesions lesions on brain MRIs. In this prospective study, 94 patients with confirmed MS (n=77) or suspected chronic inflammatory CNS disease (n=17), underwent clinically indicated brain MRI at 1.5T. Each participant underwent both c-3D-T2 (5:01 min) and DL-3D-T2 (2:48 min) imaging. Primary endpoint was interchangeability of both sequences for detecting new white matter lesions defined according to the 2024 revised McDonald criteria. Furthermore, comparability regarding total lesion count and image quality was evaluated. Lesions were assessed in 3 anatomic regions (periventricular, cortical/juxtacortical, and infratentorial) by 3 independent readers and 1 experienced neuroradiologist using an established, certified lesion-detection software. Equivalence margin for interchangeability was predefined at 5%. Gwet's AC1 and AC2 determined inter-reader reliability. The study comprised of 70 women and 24 men, with a mean age of 44.9 years (range: 24 to 72 y). In 73 patients, a comparable prior examination was available and met the technical requirements for AI-based lesion-detection software. In 13 patients new lesions were confirmed. Interchangeability for the primary endpoint was demonstrated for all readers and the experienced neuroradiologist, using certified lesion-detection software. Individual equivalence indices remained within the 5% margin. Detection of new white matter lesions demonstrated almost perfect inter-method agreement with overall Gwet's AC2 values between 0.85 [0.74; 0.96] and 0.98 [0.95; 1.00], and excellent inter-reader reliability, with overall new lesions of 0.85 [0.76; 0.94]. Total number of infratentorial and periventricular lesions demonstrated almost perfect agreement [(≥0.93 (0.91; 0.96)] and good agreement for cortical/juxtacortical lesions [0.71 (0.65; 0.77)]. Subjective analysis revealed that 2 readers rated c-3D-T2 as significantly superior in image quality (P<0.001) and 1 reader rated c-3D-T2 as significantly superior in diagnostic confidence (P=0.003). DL-accelerated 3D-T2-weighted imaging is interchangeable with conventional 3D imaging for detecting new white matter lesions, while reducing acquisition time by nearly 50%. This acceleration supports more efficient MRI protocols for routine MS surveillance, enabling inclusion of additional advanced sequences. However, as subjective image evaluation was slightly inferior in DL-accelerated data sets, and given the substantial time savings, it may be reasonable to slightly reduce the acceleration factor to enhance image quality.