Deep learning-based chemical shift-artifact correction of ZTE MRI for enhanced bone depiction of the lumbar spine.

RATIONALE AND

OBJECTIVES

Zero echo time (ZTE) is an advanced MRI technique providing CT-like images of mineralized tissues. This study evaluates the impact of deep learning (DL)-based reconstruction with chemical shift correction (DLCSC) on osseous depiction in ZTE MRI of the lumbar spine, compared to standard DL and non-DL reconstruction.

METHODS

This retrospective, single-center study included 38 patients undergoing 3 T ZTE MRI of the lumbar spine. K-space data were reconstructed using three methods: non-DL, standard DL, and a prototype DLCSC algorithm. Quantitative image sharpness, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) were analyzed using repeated-measures ANOVA. Two radiologists independently rated pathology-related criteria (n = 22) and bone depiction quality (n = 38) on a 4-point scale. Ordinal data were analyzed using the Friedman test, and inter-reader agreement was assessed with weighted Cohen's kappa.

RESULTS

DLCSC images yielded quantitatively sharper images compared to non-DL (p = 0.010) and standard DL (p < 0.001). SNR and CNR were significantly higher in DLCSC and DL than in non-DL (p < 0.001). In the qualitative assessment, mean scores for all criteria of pathologies and bone depiction quality improved significantly from non-DL and DL to DLCSC (p < 0.001). There was no evidence of differences in classification of pathologies (p > 0.05). Inter-reader agreement ranged from substantial to almost perfect (κ = 0.867-0.901).

CONCLUSION

DLCSC image reconstruction of ZTE MRI can improve bone depiction of the lumbar spine compared to standard DL and non-DL reconstruction.