Evaluation of super-resolution deep learning reconstruction on three-dimensional constructive interference in steady state for enhanced visualization of vestibular schwannomas.

Authors

Abstract

We evaluated whether, compared with conventional deep learning reconstruction (DLR) and zero-filling interpolation (ZIP), super-resolution DLR (SR-DLR) enhances the visualization of vestibular schwannomas and cranial nerves in three-dimensional constructive interference in steady state (3D-CISS) magnetic resonance (MR) imaging. This retrospective study analyzed 39 patients with vestibular schwannomas who underwent 3T MR imaging with 3D-CISS. Axial images were reconstructed using SR-DLR, DLR, and ZIP. Three readers independently evaluated tumor depiction, depiction of facial and vestibulocochlear nerves, sharpness, noise, artifacts, and diagnostic acceptability. Quantitative indicators included edge rise distance (ERD) and edge rise slope (ERS) measured along a linear region of interest (ROI) drawn perpendicularly across the tumor-cistern cerebrospinal fluid (CSF) boundary, and signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and contrast ratio (CR) calculated from circular ROIs (3-5 mm) placed in the tumor and prepontine CSF. SR-DLR significantly improved visualization of vestibular schwannoma and facial and vestibulocochlear nerves compared with ZIP (p < 0.001). Nerve depiction ratings varied between readers when comparing SR-DLR and DLR. SR-DLR yielded significantly better sharpness, lower ERD and higher ERS than both DLR and ZIP (p < 0.001). Artifacts showed no significant differences. No clear deterioration in noise was noted. Objectively, SR-DLR showed significantly higher CR than DLR (p = 0.023) and ZIP (p < 0.001). SNR and CNR were significantly higher with SR-DLR than ZIP, but comparable to DLR. SR-DLR improves sharpness and lesion conspicuity on 3D-CISS imaging without a clear increase in artifacts or noise compared with ZIP.

Tags

Topics