Leveraging a vendor-neutral deep learning reconstruction algorithm to reduce scan time and enhance image quality in T2-weighted liver MRI

Abstract This study evaluated whether a vendor-neutral deep learning reconstruction (DLR) can improve image quality in accelerated T2-weighted imaging (T2WI) of the liver MRI acquired in a single breath-hold compared with routine T2WI. In this retrospective study, 86 patients who underwent nonenhanced 3-T liver MRI in August 2024 were included. Two board-certified radiologists independently assessed T2WI obtained using three protocols: (1) routine T2WI acquired with two breath-holds (Routine-2BH); (2) accelerated T2WI reconstructed with DLR acquired in a single breath-hold (SwiftMR-1BH); and (3) accelerated T2WI reconstructed with DLR acquired with two breath-holds (SwiftMR-2BH). Quantitative signal-to-noise ratio (SNR) and qualitative image quality, including sharpness, motion artifacts, spatial mismatch, and lesion conspicuity, were evaluated. The SwiftMR protocols reduced acquisition time by approximately 9 s compared with the routine protocol. Compared with Routine-2BH protocol, both SwiftMR protocols showed significantly higher sharpness and SNR ( p < 0.001), although motion artifacts were more pronounced. Spatial mismatch scores were significantly lower with SwiftMR-1BH than with Routine-2BH and SwiftMR-2BH ( p < 0.05). Among 49 patients with focal liver lesions, lesion conspicuity was generally higher with SwiftMR protocols, although not consistently significant. Overall, vendor-neutral DLR using SwiftMR restored the image quality of accelerated single-breath-hold T2WI for liver MRI, resulting in a higher SNR and overall image quality than routine two-breath-hold T2WI.