Improved image quality and reduced acquisition time in brain MRI using deep learning-based reconstruction: A quantitative and subjective assessment compared to standard MPRAGE in 0.55 T MRI.
Authors
Affiliations (4)
Affiliations (4)
- Ruhr University Bochum, Department of Radiology, Neuroradiology and Nuclear Medicine, Johannes Wesling Hospital, Hans-Nolte-Str. 1, 32429 Minden, Germany. Electronic address: [email protected].
- Research & Clinical Translation, Magnetic Resonance, Siemens Healthineers AG, Erlangen, Germany.
- Scientific Partnerships, Siemens Healthineers AG, Erlangen, Germany.
- Ruhr University Bochum, Department of Radiology, Neuroradiology and Nuclear Medicine, Johannes Wesling Hospital, Hans-Nolte-Str. 1, 32429 Minden, Germany.
Abstract
To assess the impact of deep learning (DL)-based image reconstruction on quantitative and subjective image quality in brain MRI at 0.55 T by comparing DL-reconstructed MPRAGE with standard MPRAGE using comparable acquisition geometry and timing parameters. In this prospective study, 30 patients underwent two consecutive 3D T1-weighted MPRAGE acquisitions on a 0.55 T MRI system (MAGNETOM Free.Max, Siemens Healthineers): a standard reconstruction and a DL-based reconstruction generated from undersampled k-space data using a variational-network architecture. Identical circular regions of interest (ROIs) were placed in gray matter (caudate nucleus), white matter (centrum semiovale), cerebrospinal fluid (lateral ventricle), and air for noise estimation. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were computed. Three blinded radiologists independently evaluated image quality using a 5-point Likert scale based on anatomical detail, gray-white matter differentiation, vascular visibility, artifact burden, and overall diagnostic quality. DL-based reconstruction significantly improved objective image quality across all metrics. SNR increased from 40.83 ± 10.11 to 106.30 ± 51.38 in gray matter (p < 0.0001), from 48.95 ± 12.52 to 125.30 ± 58.01 in white matter (p < 0.0001), and from 5.77 ± 1.10 to 7.75 ± 4.76 in CSF (p = 0.021). CNR between gray and white matter increased from 8.12 ± 4.24 to 18.99 ± 11.21 (p < 0.0001). Subjective ratings favored DL reconstruction for anatomical detail (4.05 ± 0.62 vs. 3.68 ± 0.64), vascular visibility (4.37 ± 0.56 vs. 3.63 ± 0.72), and overall image quality (4.07 ± 0.72 vs. 3.57 ± 0.73). Standard reconstruction showed slightly better artifact suppression (3.80 ± 0.61 vs. 3.27 ± 0.78) and gray-white matter contrast (4.00 ± 0.64 vs. 3.87 ± 0.71). Acquisition time was reduced from 6:44 min (standard MPRAGE) to 3:06 min with DL reconstruction, corresponding to a substantial scan time reduction. DL-based reconstruction markedly enhances quantitative and subjective image quality in low-field brain MRI and enables a markedly shorter acquisition time compared to the standard protocol. This study adds clinical evidence that DL reconstruction can substantially improve 0.55 T MPRAGE imaging, supporting its integration into routine neuroimaging workflows to improve efficiency, patient comfort, and diagnostic confidence.