Deep-learning-reconstructed high-resolution 3D cervical spine MRI for foraminal stenosis evaluation

Objective To compare standard-of-care two-dimensional MRI acquisitions of the cervical spine with those from a single three-dimensional MRI acquisition, reconstructed using a deep-learning-based reconstruction algorithm. We hypothesized that the improved image quality provided by deep-learning-based reconstruction would result in improved inter-rater agreement for cervical spine foraminal stenosis compared to conventional two-dimensional acquisitions. Materials and methods Forty-one patients underwent routine cervical spine MRI with a conventional protocol comprising two-dimensional T2-weighted fast spin echo scans (2 axial planes, 1 sagittal plane), and an isotropic-resolution three-dimensional T2-weighted fast spin echo scan reconstructed over a 4-h time window with a deep-learning-based reconstruction algorithm. Three radiologists retrospectively assessed images for the degree to which motion artifact limited clinical assessment, and foraminal and central stenosis at each level. Inter-rater agreement was analyzed with weighted Fleiss's kappa (k) and comparisons between two-dimensional and three-dimensional sequences were performed with Wilcoxon signed-rank test. Results Inter-rater agreement for foraminal stenosis was "substantial" for two-dimensional sequences (k = 0.76) and "excellent" for the three-dimensional sequence (k = 0.81). Agreement was "excellent" for both sequences (k = 0.85 and 0.83) for central stenosis. The three-dimensional sequence had less perceptible motion artifact (p <= 0.001-0.036). Mean total scan time was 10.8 min for the two-dimensional sequences, and 7.3 min for the three-dimensional sequence. Conclusion Three-dimensional MRI reconstructed with a deep-learning-based algorithm provided "excellent" inter-observer agreement for foraminal and central stenosis, which was at least equivalent to standard-of-care two-dimensional imaging. Three-dimensional MRI with deep-learning-based reconstruction was less prone to motion artifact, with overall scan time savings.