Accelerated 7T 3D DESS MR Neurography of the Cervical Spine Using CS-CAIPIRINHA and Deep Learning Reconstruction.
Authors
Abstract
To evaluate the potential of combined coherent and incoherent undersampling with deep learning (DL) reconstruction for highly accelerated high-resolution three-dimensional (3D) double-echo steady-state (DESS) MRI of the cervical nerves and rootlets at 7T. In this prospective study, asymptomatic volunteers underwent 7T MRI of the cervical spine. A high-resolution, 3D DESS sequence (reconstructed voxel size 0.35 mm³ isotropic) with regular 2-fold GRAPPA undersampling served as reference. The sequence was repeated with coherent CAIPIRINHA undersampling alone and combined CAIPIRINHA + incoherent compressed sensing (CS)-based undersampling with DL reconstruction and varying acceleration factors (DL-CAIPIRINHA: R=2 to 12 and DL-CS-CAIPI: R=4 to 16), while all other imaging parameters remained constant. Two fellowship-trained musculoskeletal radiologists assessed image quality, noise, reconstruction and motion artifacts, and edge sharpness of the spinal cord, intradural rootlets/roots, and exiting spinal nerves, including dorsal root ganglia (DRGs), using 5-point Likert scales, with higher scores indicating better image quality or fewer artifacts. Statistics included Friedman and post hoc Wilcoxon signed-rank tests (Benjamini-Hochberg corrected) and κ statistics. Thirty-two volunteers (mean age 30.9±7.2 y; mean BMI 23.4±2.8 kg/m²; 16 females) were included. Image quality significantly improved with 2- and 4-fold DL-CAIPIRINHA (means 4.4 to 4.6) and 4- and 8-fold DL-CS-CAIPI (3.8 to 3.9) compared with GRAPPA (3.2; all P<0.001), and remained comparable with 12-fold DL-CS-CAIPI (P=0.42) despite more than 6-fold reduction in scan time. All DL-based reconstructions reduced noise (P<0.001). Reconstruction artifacts were more pronounced for 12-fold DL-CAIPIRINHA (1.6) than for 12- and 16-fold DL-CS-CAIPI (3.1 to 3.3; P<0.001). Motion artifacts decreased with acceleration factors ≥4 (P<0.001). Sharpness of the spinal cord, intraspinal rootlets/roots, and DRGs was highest with 4-fold DL-CAIPIRINHA (4.2 to 4.7) vs. GRAPPA (2.9 to 3.2; P<0.001). Inter-reader agreement was almost perfect across all imaging features (κ=0.85 to 0.95). Coherent CAIPIRINHA undersampling with DL reconstruction improves image quality over conventional GRAPPA while enabling 50% faster imaging in 7T cervical spine MRI. Adding incoherent CS undersampling permits substantially higher acceleration with preserved image quality and fewer reconstruction artifacts. The novel integration of coherent and incoherent undersampling via combined CAIPIRINHA and CS with DL reconstruction provides a robust framework for preserving image quality at acceleration factors beyond 8, enabling efficient high-resolution cervical spine MRI with scan time reductions exceeding 6-fold.