A retrospective cohort study was performed in accordance with the Declaration of Helsinki and the STROBE guidelines. The need for ethical approval was waived. If required by regulations of participating centers, patients not occurring in objection registers were sent a letter detailing study procedures and aims and could opt out of the study. All patients with pathology-proven FNH, HCA, or HCC (biopsies or resection specimens) between 2010 and 2020 were identified through local pathology databases. Radiologists in participating centers reviewed all Gd-EOB-DTPA-enhanced MRIs. All lesions hyper- or isointense compared to the surrounding liver in the HBP by visual interpretation were included, including partially hyper- to isointense lesions.

All hyper- or isointense lesions were centrally pseudonymously reviewed by two expert radiologists (more than twenty years of experience) blinded from the pathological classification. Included were patients with hyper- or isointense regions in the lesions compared to the surrounding liver, not attributable to pre-contrast signal intensity, therefore due to contrast uptake in the HBP. In case of doubt, subtraction images were produced.

In general, MRI scans were conducted using 1.5-Tesla or 3.0-Tesla scanners. Typically, a spoiled 3D gradient fat-saturated T1-weighted sequence was produced before and during the different dynamic phases after contrast injection. The HBP consisted of a similar sequence as the dynamic sequence, and was performed between 10 min and 20 min after contrast injection. In- and opposed-phase T1-weighted sequence was performed using the Dixon technique, if available. T2-weighted 2D spin-echo sequence was performed with or without fat saturation, using a single shot or/and multishot protocol. Diffusion-weighted sequences included typically minimally two b-values with the first between 0 s/mm² and 150 s/mm² and the second from 600 s/mm² to 800 s/mm².

The supplementary material shows full signs, definitions, and scoring process. Patterns in the HBP were raster, perilesional hypointense rim [25], lesional hyper- or isointense rim [26], atoll fingerprint, white-bordered flower (present/absent), partification (no, craquelure with loosening part, focal/mosaic or nodule-in-nodule) [27, 28], and extent of hyper- or isointensity (limited/extensive). Patterns on T2-w were sinusoidal dilatation (present/absent) [23] and partification (no, focal/mosaic or nodule-in-nodule). Patterns scored in the HBP and T2-weighted images were scar [29], liquid (present/absent), and extent of heterogeneity (no/limited or extensive).

Miscellaneous imaging features were: T1-w in-phase hyperintensity not due to fat, hemosiderin (present/absent), and intralesional fat (absent, less than 50%, or more than 50%). In dynamic phases, intralesional arterial hypointense rim (present/absent) and venous washout (present, absent, equivocal, or pseudo-washout) were scored [9, 30, 31]. If diffusion-weighted imaging (DWI) was available, diffusion restriction was scored (present, absent, or equivocal). Lesion contour was defined as spherical, cauliflower, or aspecific.

Finally, the final diagnosis (i.e., FNH or HCA/HCC) and treatment recommendation (i.e., no follow-up, follow-up, or histology (either biopsy or resection)) were formulated. HCA and HCC were grouped because both are hepatocellular lesions with (risk of) malignant transformation generally assumed to be hypointense in the HBP. Treatment recommendations were based on imaging characteristics alone, not taking lesion size and location into account.

Pathological classification was based on a review of pathology reports. The time between MRI imaging and biopsy and/or resection was a median of 0 months (IQR: 0–3 months). The electronic supplementary material describes morphological requirements for each diagnosis. For the differentiation between FNH and HCA, either a pattern of glutamine synthetase (GS) staining or, in the case of HCA, a positive C-reactive protein or positive serum amyloid A staining was required [32]. B-catenin-activated HCA was characterized by nuclear staining of B-catenin and/or overexpression of GS (n = 5) and/or molecular evidence of mutations in the CTNNB1 gene (n = 9). If available, the locus of the mutation was noted.

Clinical data and baseline characteristics collected were lesion size in centimeters and lesion location (left or right liver lobe). Pathological and clinical data were collected using RedCap. Underlying liver disease was defined as non-alcoholic fatty liver disease or non-alcohol steatohepatitis, cirrhosis and/or fibrosis, and metabolic diseases including hepatocyte nuclear factor 1A maturity-onset diabetes of the young and glycogen storage disorders.

Statistical analyses were conducted using SPSS® statistics version 28.0 (IBM). Classification and regression tree (CART) analysis was conducted using the recursive partitioning and regression trees (Rpart) package in RStudio version 4.2.1 (Posit). The Rpart package uses ten-fold cross-validation to determine the optimal place to prune the regression tree. Categorical variables were expressed as numerators/denominators with percentages and continuous variables as means with standard deviations (SDs) or medians with interquartile ranges (IQRs). For all analyses, available case analysis was conducted.

Interobserver and intraobserver variability were calculated using Cohen’s kappa statistic. Levels of agreement were interpreted using cut-off values defined by Landis and Koch: < 0.00 poor, 0.00–0.20 slight, 0.21–0.40 fair, 0.41–0.60 moderate, 0.61–0.80 substantial, and 0.81–1.00 almost perfect [33].

In univariable analyses, imaging features were compared between patients with a final pathological diagnosis of FNH and HCA/HCC using Pearson’s chi-squared test or Fisher’s exact test. p-values < 0.05 were considered statistically significant. Diagnostic odds ratios with 95% confidence intervals for the diagnosis FNH, compared to HCA/HCC were calculated. Sensitivity, specificity, positive predictive value, and negative predictive value were calculated for significant imaging features. Haldane–Anscombe correction was performed in the case of zero cells.

For multivariable analyses, imaging features with a p-value of > 0.25 in univariable analysis were excluded. A Spearman’s correlation matrix of imaging features was derived. In cases of a correlation coefficient of 0.80 and higher, indicative of collinearity, one of the correlated imaging features was excluded from the multivariable analysis. The tree diagram obtained by CART analysis and odds ratios obtained by multivariable logistic regression were combined in a visual representation made in the draw.io® (JGraph Limited). Sensitivity analyses assessed the occurrence of imaging features suggestive of FNH in B-catenin-activated HCA.