An Inline-Integrable Image Restoration Model With ROI Fine-Tuning and Slice-to-Volume Reconstruction for Accelerated Whole-Brain Vessel Wall Imaging.

PURPOSE: To develop a vessel wall imaging (VWI)-dedicated image restoration model enabling high-resolution whole-brain VWI in under 6 min. METHODS: A transformer-based 2.5D model was trained in a supervised manner to enhance the image quality of CAIPIRINHA 2 × 2-accelerated VWI acquisitions to a level comparable to that of GRAPPA twofold-accelerated scans. The majority of k-space data used for training and testing were acquired on a Siemens 3 T system equipped with a standard 20-channel head-neck coil. Region-of-interest fine-tuning and slice-to-volume reconstruction were devised to balance global structural fidelity with local vascular precision and 3D continuity. Model performance was evaluated both quantitatively and qualitatively through single-center retrospective testing and prospective testing in multi-center, generalized settings. Integration of the model into the vendor's inline image reconstruction pipeline was also investigated. RESULTS: Compared with the input images, model outputs improved peak SNR by more than 7 dB and structural similarity index measure by 0.19 in the retrospective test set. When applied to CAIPIRINHA fourfold-accelerated data that otherwise underwent offline or inline GRAPPA reconstruction, the model significantly enhanced vessel wall SNR, wall-to-lumen CNR, wall-to-cerebrospinal fluid CNR, and image quality scores. These improvements were sustained in an independent cohort from different institutions and reproduced in a dataset acquired with smaller field of view and a 64-channel coil, supporting the model's robustness and generalizability. CONCLUSION: The proposed inline-integrable image restoration model enables clinically viable whole-brain VWI in under 6 min using standard 20-channel coils, offering a practical approach to improve patient throughput without compromising diagnostic quality.