Clinical evaluation of deep learning-accelerated 3D FLAIR imaging at 1.5T and 3T for visualization of cerebral white matter lesions.
Authors
Affiliations (12)
Affiliations (12)
- Department of Radiology, Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, USA.
- Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA.
- Department of Radiology, Faculty of Medicine, Siriraj Hospital, Mahidol University, Bangkok, Thailand.
- Department of Neurology, University of New England College of Osteopathic Medicine, Biddeford, ME, USA.
- Department of Medical Imaging, Shuang Ho Hospital, Taipei Medical University, New Taipei City, Taiwan.
- Department of Radiology, School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan.
- GE HealthCare, Boston, MA, USA.
- GE HealthCare, Rochester, MN, USA.
- Department of Neurology, University of Massachusetts Memorial Medical Center and University of Massachusetts Chan Medical School, Worcester, MA, USA.
- National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, MD, USA.
- Enterprise Service Group - Radiology at Wentworth-Douglass Hospital, Mass General Brigham, Dover, NH, USA.
- Department of Radiology, University of Turku, Turku, Finland.
Abstract
BackgroundFluid-attenuated inversion recovery (FLAIR) imaging is a standard sequence used for the detection of white matter lesions; however, 1-mm isotropic FLAIR can take 5-7 min at 1.5T and 3T using routine methods, limiting routine use of high-resolution 3D FLAIR imaging.PurposeTo evaluate a deep learning (DL)-based unrolled optimization 3D FLAIR technique, Sonic DL™ 3D (SDL), in comparison to a standard 3D FLAIR protocol, both combined with DL-based post-processing (AIR™ Recon DL) at 1.5T and 3T.Material and MethodsA total of 67 patients (15 scanned at 1.5T and 52 at 3T) with suspected white matter lesions were included. Two neuroradiologists performed blinded, head-to-head image quality evaluation (motion, noise, and overall diagnostic quality). Quantitative lesion volume and lesion counts were derived from automated segmentation using SPM 12. Cohen's kappa, Wilcoxon signed-rank tests, and paired <i>t</i>-tests were applied.ResultsAt 1.5T, SDL FLAIR (4 min 53 s) reduced scan times by 15% compared to standard 3D FLAIR (5 min 43 s) and performed similarly or better in lesion visualization and image quality, with significantly (<i>P</i> <0.05) improved noise quality. At 3T, SDL FLAIR (2 min 45 s) achieved a reduction in scan time of approximately 35% compared to standard 3D FLAIR (4 min 15 s) and showed improved lesion conspicuity (periventricular, juxtacortical, and other regions) and image quality metrics. However, except for noise quality, these differences did not reach statistical significance despite moderate to almost perfect interrater agreement (Cohen's kappa 0.55-0.85; <i>P</i> <0.001).ConclusionAccelerated SDL FLAIR demonstrated improved diagnostic quality, comparable lesion conspicuity, and volumetric measurements to standard 3D FLAIR sequence at 1.5T and 3T.