Super‐resolution reconstruction of MR vessel wall images using a deep neural network

Abstract Background Magnetic resonance vessel wall imaging (MR‐VWI) is a non‐invasive, high‐resolution technique that enables detailed visualization of vascular structures and plays a crucial role in diagnosing cerebrovascular diseases. However, imaging intracranial perforating arteries requires exceptionally high spatial resolution. Conventional super‐resolution techniques primarily focus on real‐valued reconstructions, neglecting the inherently complex‐valued nature of MR data and failing to fully exploit phase information. Since MR images are inherently acquired in complex form, relying solely on magnitude data discards valuable phase components essential for comprehensive vascular assessment. Purpose To develop and evaluate a deep‐learning‐based framework capable of generating high‐resolution MR vessel wall images from low‐resolution acquisitions by modeling the complex‐valued nature of MR data. Methods A three‐dimensional complex‐valued super‐resolution (CVSR) neural network was designed to reconstruct high‐resolution images while preserving both magnitude and phase information within the complex domain. The CVSR model was trained on 200 data sets and tested on 50 pairs. Three super‐resolution approaches, including Fourier interpolation, an enhanced deep super‐resolution (EDSR) network with two real‐valued input channels (EDSR‐2Ch), and the proposed CVSR, were compared against ground‐truth images at 0.44 mm 3 isotropic resolution using structural similarity (SSIM), peak signal‐to‐noise ratio (PSNR), and root‐mean‐square error (RMSE). Results The proposed CVSR achieved superior quantitative performance and visual fidelity compared with competing methods. It reconstructed vessel wall images with improved clarity, continuity, and closer resemblance to the ground truth. On the 50‐subject test set, CVSR achieved higher SSIM (0.771 vs. 0.759 and 0.628), PSNR gains of 0.35 and 2.58 dB, and RMSE reductions of 3.97% and 25.77% compared with EDSR‐2Ch and Fourier interpolation, respectively. Conclusions The proposed CVSR framework effectively transforms low‐resolution MR vessel‐wall images into high‐resolution reconstructions, enhancing detail visualization and delineation of fine vascular details. This approach has the potential to improve the clinical assessment of cerebrovascular pathology, particularly arterial wall and plaque characterization.