Phase-adaptive automated whole-liver quantification of signal intensity-derived liver indices from gadoxetic acid-enhanced MRI.

objectivesTo develop and validate a phase-adaptive pipeline for automated whole-liver, parenchyma-based quantification of signal intensity-derived liver indices from gadoxetic acid-enhanced MRI without inter-phase co-registration. MATERIALS AND

methodsIn this retrospective, single-center study (2016-2025), 9,013 liver MRI examinations were screened, and 3,112 gadoxetic acid-enhanced MRI examinations were identified. The first examination per patient defined the clinical cohort (1,174 patients; mean age, 59.6 ± 14 years [range, 4-93 years]; 611 men). A phase-adaptive nnU-Net was trained on 403 and tested on 500 dynamic contrast-enhanced MRI sequences for segmentation of liver parenchyma, intrahepatic non-parenchymal structures, and spleen; paraspinal muscle reference regions were obtained using TotalSegmentator. Segmentation accuracy was assessed against manually corrected reference segmentations generated by a semi-automatic annotation workflow with Dice similarity coefficients. Automated arterial enhancement fraction (AEF), relative liver enhancement (RLE), hepatocellular uptake index (HUI), liver-to-spleen ratio (LSR), and liver-to-muscle ratio (LMR) were compared with manual region-of-interest measurements in a radiological validation cohort (50 patients) using intraclass correlation coefficients and Bland-Altman analysis. Liver indices were compared between patients with cirrhotic and non-cirrhotic livers using the Mann-Whitney U-test.

resultsLiver parenchyma segmentation achieved Dice similarity coefficients of 0.998-0.999. Reader-model agreement for the liver indices ranged from 0.936 to 0.977. HUI, LSR, RLE, and LMR were lower in cirrhosis (p < 0.001), whereas AEF did not differ (p = 0.14).

conclusionPhase-adaptive segmentation enables whole-liver quantification of signal intensity-derived liver indices without inter-phase co-registration. The lower hepatobiliary uptake indices in cirrhosis support the plausibility of this approach for scalable quantitative assessment of liver MRI. KEY POINTS: Question Can gadoxetic acid-enhanced liver MRI be quantified automatically across contrast phases without inter-phase co-registration or manual region-of-interest placement in routine clinical imaging? Findings Phase-adaptive segmentation enabled automated whole-liver quantification with liver Dice coefficients of 0.998-0.999 and excellent agreement with manual measurements for five liver indices. Relevance Statement Automated whole-liver quantification from standard gadoxetic acid-enhanced liver MRI provides reproducible, signal-intensity-derived liver indices and reduces dependence on manual region-of-interest measurements in large-scale clinical studies.