Abstract:Objectives: Deep learning-reconstructed diffusion-weighted imaging (DL-DWI) is an emerging promising time-efficient method for liver evaluation, but analyses regarding different motion compensation strategies are lacking. This study evaluated the qualitative and quantitative features, sensitivity for focal lesion detection, and scan time of free-breathing (FB) DL-DWI and respiratory-triggered (RT) DL-DWI compared with RT conventional DWI (C-DWI) in the liver and a phantom. Materials and methods: Eighty-six patients indicated for liver MRI underwent RT C-DWI, FB DL-DWI, and RT DL-DWI with matching imaging parameters other than the parallel imaging factor and number of averages. Two abdominal radiologists independently assessed qualitative features (structural sharpness, image noise, artifacts, and overall image quality) using a 5-point scale. The signal-to-noise ratio (SNR) along with the apparent diffusion coefficient (ADC) value and its standard deviation (SD) were measured in the liver parenchyma and a dedicated diffusion phantom. For focal lesions, per-lesion sensitivity, conspicuity score, SNR, and ADC value were evaluated. Wilcoxon signed rank test and repeated-measures analysis of variance with post hoc test revealed the difference in DWI sequences. Results: Compared with RT C-DWI, the scan times for FB DL-DWI and RT DL-DWI were reduced by 61.5% and 23.9%, respectively, with statistically significant differences between all 3 pairs (all P 's < 0.001). Respiratory-triggered DL-DWI showed a significantly sharper liver margin, less image noise, and more minor cardiac motion artifact compared with RT C-DWI (all P 's < 0.001), whereas FB DL-DWI showed more blurred liver margins and poorer intrahepatic vessels demarcation than RT C-DWI. Both FB- and RT DL-DWI showed significantly higher SNRs than RT C-DWI in all liver segments (all P 's < 0.001). There was no significant difference in overall ADC values across DWI sequences in the patient or phantom, with the highest value recorded in the left liver dome by RT C-DWI. The overall SD was significantly lower with FB DL-DWI and RT DL-DWI than RT C-DWI (all P 's ≤ 0.003). Respiratory-triggered DL-DWI showed a similar per-lesion sensitivity (0.96; 95% confidence interval, 0.90-0.99) and conspicuity score to those of RT C-DWI and significantly higher SNR and contrast-to-noise ratio values ( P ≤ 0.006). The per-lesion sensitivity of FB DL-DWI (0.91; 95% confidence interval, 0.85-0.95) was significantly lower than that of RT C-DWI ( P = 0.001), with a significantly lower conspicuity score. Conclusions: Compared with RT C-DWI, RT DL-DWI demonstrated superior SNR, comparable sensitivity for focal hepatic lesions, and reduced acquisition time, making it a suitable alternative to RT C-DWI. Despite FB DL-DWI's weakness in motion-related challenges, further refinement could potentiate FB DL-DWI in the context of abbreviated screening protocols, where time efficiency is a high priority.

Deep Learning-Accelerated Liver Diffusion-Weighted Imaging: Intraindividual Comparison and Additional Phantom Study of Free-Breathing and Respiratory-Triggering Acquisitions