Quantification of edge-enhancing effects using accelerated deep learning reconstructed orbital MRI sequences

To qualitatively and quantitatively compare image quality, edge sharpness, and internal-structure delineation of accelerated deep learning-reconstructed turbo spin-echo (TSE_DLR) sequences versus conventionally reconstructed TSE (TSE_CR) in 3 T orbital MRI.

Twenty-five patients undergoing 3 T orbital MRI were retrospectively selected. All underwent conventional and accelerated (up to 75%) deep learning-reconstructed axial and coronal T2-weighted and contrast-enhanced fat-saturated T1-weighted TSE (T1CEfs) sequences. Two blinded neuroradiologists rated image quality, edge sharpness, internal tissue delineation, overall sharpness, lesion detection confidence, and lesion conspicuity on a 5-point Likert scale. Quantitative analyses included similarity and image-quality metrics, the Perceptual Sharpness Index (PSI), blur metrics, and edge parameters after Roberts, Sobel, or Canny edge extraction.

Qualitatively, edge and overall sharpness were superior with TSE_DLR in T1CEfs (all p < 0.001), and mainly in T2-weighted imaging. Internal tissue delineation was rater-dependent, with the experienced rater favoring TSE_CR. Lesion-detection confidence did not differ significantly across reconstructions. Quantitatively, TSE_DLR improved edge sharpness, with increased edge steepness (all p < 0.001) and decreased edge width (most p ≤ 0.003), though edge contrast was slightly reduced. Blur metrics favored TSE_CR, while PSI (all p < 0.001) and SNR/PSNR favored TSE_DLR. Structural similarity was high (SSIM 0.986-0.995; MS-SSIM 0.998-0.999), while high-frequency similarity was lower (FSIM: 0.804-0.861; WASH: 0.729-0.747).

Accelerated DL-based reconstructions for orbital MRI maintain or improve lesion conspicuity, overall image quality, and edge sharpness, while substantially reducing scan time. However, reduced internal anatomical delineation warrants cautious interpretation of the images.