Super-Resolution Deep Learning Reconstruction for 3-Dimensional T1-Weighted Gradient-Echo Imaging in Contrast-Enhanced Liver MRI: Comparison With Conventional and Standard Deep Learning Reconstructions.

Purpose

To evaluate the image quality of super-resolution deep learning reconstruction (SR-DLR) for 3-dimensional (3D) T1-weighted gradient-echo (GRE) imaging in contrast-enhanced MRI, compared with conventional reconstruction (Conv.) and standard deep learning reconstruction (DLR).

Materials And Methods

This retrospective study included 50 patients (mean age: 71.5 y) with 76 hepatic lesions who underwent contrast-enhanced dynamic liver MRI at 3T. Portal venous phase images were reconstructed using Conv., DLR, and SR-DLR. Quantitative analyses measured liver signal-to-noise ratio (SNR), contrast ratio (CR) between liver parenchyma and lesions, edge rise distance (ERD), and edge rise slope (ERS). Qualitative assessments of image noise, sharpness, contrast, motion artifacts, overall image quality, and lesion conspicuity were performed independently by 2 radiologists using a 5-point scale. Coronal multiplanar reformatted (MPR) images were also evaluated. Statistical comparisons were performed using the Friedman test with Bonferroni correction.

Results

Liver SNR was comparable between SR-DLR and DLR. SR-DLR produced sharper edges (lower ERD, higher ERS) and higher lesion-to-liver CR than Conv. and DLR (P<0.001). The overall CR differences were primarily driven by cystic lesions, whereas improvements in malignant lesions were modest. Qualitatively, SR-DLR received the highest ratings for sharpness, overall image quality, and lesion conspicuity, particularly on MPR images (P<0.001). SR-DLR images exhibited slightly higher image noise and motion-related artifacts than DLR.

Conclusion

SR-DLR improved spatial resolution, sharpness, and lesion conspicuity in contrast-enhanced MRI compared with Conv. and DLR, indicating potential to improve the diagnostic performance of dynamic liver MRI.