Reducing breath-hold time in liver MRI: Clinical performance of deep learning-accelerated post-contrast T1 VIBE...

Post-contrast liver MRI often requires long breath-holds, risking motion artifacts that can reduce diagnostic quality. We assessed whether deep learning-accelerated acquisitions can shorten breath-holds while maintaining diagnostic image quality.

In this prospective study, patients underwent three post-contrast venous-phase T1-weighted two-point Dixon gradient-echo sequences on a 1.5 T system: a standard (18 s acquisition time) and two deep learning-accelerated sequences (10 s and 6 s). Three blinded radiologists rated overall image quality, motion artifacts, other artifacts, anatomic differentiability, and lesion conspicuity on 5-point Likert scales. Per-patient consensus was defined as the median across readers. Within-patient differences were tested with the Friedman test and Holm-adjusted Wilcoxon signed-rank tests. Non-inferiority of diagnostic acceptability (defined as Likert ≥ 3) was tested using a prespecified margin of - 5% compared with the standard sequence.

We enrolled 99 patients (mean age 61.0 ± 15.4 years, 49.5% females). The standard sequence had higher ratings for anatomic differentiability (5 vs. 4 and 4), and lesion conspicuity (5 vs. 4 and 4¸ both p < 0.001) and modestly higher overall image quality despite identical medians (4 vs. 4 and 4, p < 0.001). Motion artifact ratings did not differ across sequences. No inferiority was noted for both accelerated sequences for all items.

Deep learning-based reconstruction enabled substantial acceleration of post-contrast liver MRI, reducing breath-hold time by up to 67%. Despite minor image quality trade-offs, diagnostic value remained acceptable, supporting motion-robust, faster post-contrast liver MRI.