Deep-Learning Accelerated Vessel Wall Imaging Using T1-SPACE at Ultra-High-Field Strength MRI.

BACKGROUND AND PURPOSE: Current literature on deep learning-accelerated intracranial vessel wall imaging has been largely limited to postprocessing-based approaches at 3T, with comparatively sparse data at 7T. The purpose of the present study was to evaluate the feasibility and technical performance of a postcontrast T1-SPACE sequence using deep learning-based image reconstruction (DLBIR) for intracranial vessel wall imaging (IC-VWI) at 7T, comparing image quality, artifacts, and acquisition time with the standard-of-care (SOC) T1-SPACE sequence.

MATERIALS AND METHODS: In this retrospective single-center study, 36 patients (21 women; mean age, 53.3 ± 16.2 years) underwent IC-VWI at 7T using both SOC T1-SPACE and DLBIR-accelerated T1-SPACE sequences. Two independent neuroradiologists assessed overall image quality (noise, artifacts, sharpness, and overall quality) and wall and lumen visualization along the intracranial vessels using a 4-point Likert scale. A cumulative-logit mixed-effects model (CLMM) of segment ratings was used for intersequence comparison. Segments were also pooled into proximal versus distal vessels and quality was compared using the exact paired sign test. Overall image-quality metrics were compared using paired Wilcoxon tests and Bland-Altman plots. Interreader agreement was summarized using percent agreement.

RESULTS: On segment-level analysis, T1-SPACEDL yielded markedly higher ratings than T1-SPACESOC (wall: OR: 22.79 [95% CI: 15.83-32.82]; lumen: OR: 97.00 [95% CI: 66.80-141.00]; both false discovery rate (FDR)-adjusted P < .001). Effects remained large when segments were pooled into proximal and distal segments (for wall, proximal segment OR: 297.00 [95% CI: 40.00-∞] and distal segment OR: 71.00 [28.00-404.00]; for lumen, proximal OR: 82.00 [36.00-339.00] and distal OR: 225.00 [60.00-13350.00]; all FDR-adjusted P < .001)]. Reader-wise Wilcoxon tests showed higher overall image-quality scores across noise, artifacts, sharpness, and overall quality for T1-SPACEDL (all P < .001). Interreader agreement was uniformly high with T1-SPACEDL when compared with T1-SPACESOC. Bland-Altman analysis demonstrated a positive bias favoring T1-SPACEDL. Acquisition time was reduced from 7:30 minutes with T1-SPACESOC to 6:00 minutes (20% reduction) with T1-SPACEDL.

CONCLUSIONS: Contrast-enhanced T1-SPACEDL at 7T may have potential clinical utility for IC-VWI, given the shorter Acquisition time, improved image quality and reduced image artifacts.