Optimizing high-resolution knee MRI at 3 tesla: conventional acceleration versus deep learning reconstruction
摘要
To compare subjective and objective image quality and resolution between time-optimized standard knee MRI (sMRI) with image quality-optimized DL-enhanced (DL-MRI) at 3 Tesla.
MethodsA retrospective single-centre study of 150 knee MRI examinations (75 sMRI, 75 DL-MRI) was conducted. Protocols included Proton density–weighted sequence with fat suppression (PD-FS) (coronal/sagittal/axial), T1 (coronal or sagittal), and T2 (axial) optimized for time in sMRI and for image quality in DL-MRI. Three blinded readers with different levels of experience rated overall image quality, anatomical delineation, fat saturation, motion artefacts, and foreign-body artefacts on 5-point Likert scales. Quantitative analysis was performed to calculate SNR, CNR, and generalised metrics (gSNR, gCNR). Group differences were assessed using two-sided Welch’s t-tests.
ResultsReaders rated DL higher in nearly all categories and sequences, with mean gains of ~ 0.48–0.70 for overall image quality and ~ 0.38–0.54 for anatomical delineation (all p ≤ 0.001). Fat saturation improved for PD-FS coronal and axial, motion artefacts improved for PD-FS coronal and sagittal (and slightly for axial), and foreign-body artefacts were comparable. Quantitatively, PD-FS showed higher muscle SNR and higher gSNR/gCNR with DL (bone SNR non-significant); T2 showed higher bone SNR and higher CNR/gCNR but lower muscle SNR/gSNR; and T1 showed lower SNR/gSNR with preserved CNR. Compared with sMRI, DL-MRI achieved a twofold improvement in in-plane resolution (0.4 × 0.4 mm² vs. 0.2 × 0.2 mm²) and reduced slice thickness (3.0 mm vs. 2.5 mm. and down to 1.0 mm for T2-weighted sequences), while slightly shortening total scan time (10:14 min vs. 9:30 min).
ConclusionDL-MRI provided superior image quality and higher resolution over time-optimized standard knee MRI at 3 Tesla.
Trial registrationNot applicable.