Enhanced Neurovascular Imaging Using Ultra-High-Resolution CT and Deep Learning-Based Image Reconstruction.
Authors
Affiliations (5)
Affiliations (5)
- From the Department of Neuroradiology (S.S., M.A.A.M., M.F., A.K., M.A.B., A.E.O.), University Medical Center Mainz, Johannes Gutenberg University, Mainz, Germany [email protected].
- From the Department of Neuroradiology (S.S., M.A.A.M., M.F., A.K., M.A.B., A.E.O.), University Medical Center Mainz, Johannes Gutenberg University, Mainz, Germany.
- Technical University of Darmstadt (A.S.), Darmstadt, Germany.
- Canon Medical Systems Corporation, CT-MR Division (K.H.), Kanagawa, Japan.
- Department of Neurology (M.H., T.U.), University Medical Center Mainz, Johannes Gutenberg University, Mainz, Germany.
Abstract
CTA is an established technique for imaging intracranial arteries, enabling rapid assessment of stenosis, vessel occlusions and aneurysm in various acute and elective clinical settings. This study aims to evaluate. Our purpose was to evaluate the diagnostic benefits of deep learning-based image reconstruction for neurovascular imaging by using ultra-high-resolution (UHR)-CT compared with standard hybrid iterative reconstruction (HIR) applied to both UHR-CT and normal-resolution (NR)-CT data sets. This retrospective, single-center study included 100 consecutive patients who underwent cranial CT and CTA for acute neurologic symptoms. Imaging was performed on a UHR-CT system. HIR was applied to CTA data sets reconstructed with: 1) an NR matrix of 512 × 512 pixels and 0.5 mm slice thickness (NR-CTA) and 2) a UHR matrix of 1024 × 1024 pixels and 0.25 mm slice thickness (UHR-CTA). Downscaling from the UHR data was performed for the NR-CTA by averaging 4 voxels (2 × 2) into 1 voxel, effectively converting the 1024 × 1024 matrix to a 512 × 512 matrix. A vendor-specific deep-learning algorithm trained for neurovascular analysis was additionally applied to UHR-CTA data sets (deep-learning UHR-CTA [DL-UHR-CTA]). Quantitative analyses included SNR, contrast-to-noise ratio (CNR), and slope evaluations in 3 vessel sections: the MCA, basilar artery (BA), and a subcortical vessel (SV) by using a Matlab-tool. Qualitative assessments of image quality, contrast, artifacts, diagnostic confidence, and vessel assessability (proximal, intermediate, and subcortical segments) were conducted by 2 radiologists by using a 4-point Likert scale. No significant differences between DL-UHR-CTA and NR-CTA were observed in SNR and CNR for BA and MCA; however, DL-UHR-CTA outperformed NR-CTA in SNR and CNR for SV (<i>P</i> < .001). NR-CTA revealed significantly lower SNR for BA and MCA (<i>P</i> < .05) and lower CNR for BA (<i>P</i> = .02) compared with UHR-CTA. No significant differences in CNR for MCA and SNR for SV were observed between NR-CTA and UHR-CTA. DL-UHR-CTA (-359.6 ± 116.0) was significantly steeper than both NR-CTA (-226.5 ± 64.2 Hounsfield unit [HU]) and UHR-CTA (-249.2 ± 67.1 HU) across all vessel segments (<i>P</i> < .001). Qualitative analysis showed DL-UHR-CTA provided significantly better overall image quality, contrast, diagnostic confidence, and accessibility across all vessel segments, with fewer artifacts, compared with UHR-CTA and NR-CTA (<i>P</i> < .05). Deep learning-based reconstruction of UHR-CTA images in neurovascular imaging significantly improves overall image quality, vascular delineation, SNR, and CNR compared with HIR alone.