Image Quality of 1.5 T MRI Using Dual-Type Deep Learning Reconstruction for Pancreatic Cystic Lesions: An Exploratory Comparison With Conventional 3 T MRI Without Deep Learning Reconstruction.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, National Defense Medical College, Tokorozawa, Saitama, Japan.
- Center for Trauma, Burn and Tactical Medicine, National Defense Medical College Hospital, Tokorozawa, Saitama, Japan.
Abstract
Dual-type deep learning reconstruction (DLR) may improve the pancreatic MRI image quality; however, the effects of different DLR strengths at 1.5 T compared with the conventional 1.5 T and 3 T MRI without DLR remain unclear. To evaluate the image quality of 1.5 T MRI reconstructed with different DLR strengths for pancreatic cystic lesions, compared with conventional 1.5 T and 3 T MRI without DLR. Retrospective, single-center study. Ninety-nine patients with pancreatic cystic lesion (53 men, 46 women; age range, 29.2-93.8 years) underwent 1.5 T MRI; 54 had undergone 3 T MRI within the preceding 2 years. 1.5 T and 3 T MRI; axial single-shot turbo spin-echo T2-weighted imaging and heavily T2-weighted single-shot turbo spin-echo two-dimensional MRCP. The 1.5 T images were reconstructed using dual-type DLR with 80% and 50% denoising levels and compressed sensing (CS) without DLR. The 3 T images used CS without DLR. Three blinded radiologists independently assessed eight image-quality items using a 5-point Likert scale. Signal-to-noise ratio, contrast-to-noise ratio, and edge rise distance were measured. Friedman tests followed by Wilcoxon signed-rank tests with Holm correction were used. At 1.5 T, 80% DLR yielded significantly higher scores than CS for all qualitative items except for texture naturalness with higher SNR (40.20 vs. 23.23) and CNR (24.67 vs. 13.57) and lower ERD (1.59 vs. 1.98 mm). ERD did not differ between 80% and 50% DLR (p = 0.825). Compared with 3 T CS, 1.5 T 80% DLR showed significantly higher scores for seven of eight items, significantly higher SNR and CNR, and lower ERD (39.69 vs. 30.62; CNR, 24.33 vs. 16.17; and 1.49 vs. 2.16 mm, respectively). Texture naturalness was significantly lower with 80% DLR, indicating a trade-off between image enhancement and naturalness. In this exploratory, non-simultaneous comparison, 1.5 T DLR showed favorable image-quality metrics relative to 3 T CS without DLR. 3. Stage 1.