Inter-observer agreement and accuracy of LI-RADS v2018 for differentiating tumor in vein from bland thrombus using gadoxetic acid-enhanced magnetic resonance imaging.

PURPOSE: To assess inter-observer agreement and accuracy of LI-RADS v2018 for differentiating tumor in vein (TIV) from bland thrombus on gadoxetic acid-enhanced magnetic resonance imaging (Gx-MRI). Secondarily, to determine whether a multi-feature model improves accuracy compared to LI-RADS.

METHODS: We retrospectively identified consecutive patients at risk for hepatocellular carcinoma with venous occlusion(s) reported on Gx-MRI. Five radiologists independently classified each occlusion as TIV or bland thrombus using the LI-RADS TIV criterion (enhancing soft tissue in vein). They also evaluated imaging features suggestive of TIV or bland thrombus. Intra-class correlation coefficient (ICC) was calculated for individual features. A multi-feature model was developed based on consensus scores of features with > 5% consensus prevalence and > 0.40 ICC. Sensitivity and specificity of the LI-RADS criterion and of the cross-validated multi-feature model were compared.

RESULTS: Ninety-eight patients with 103 venous occlusions (58 TIV, 45 bland thrombus) were included. The LI-RADS criterion provided 0.63 ICC and, depending on the reader, 0.62-0.93 sensitivity and 0.87-1.00 specificity. Five other features had > 5% consensus prevalence and > 0.40 ICC, including three LI-RADS suggestive features and two non-LI-RADS features. The optimal multi-feature model incorporated the LI-RADS criterion and one LI-RADS suggestive feature (occluded or obscured vein contiguous with malignant parenchymal mass). After cross-validation, the multi-feature model did not improve sensitivity or specificity compared to the LI-RADS criterion (P = 0.23 and 0.25, respectively).

CONCLUSION: Using Gx-MRI, the LI-RADS criterion for TIV provides substantial inter-observer agreement, variable sensitivity, and high specificity for differentiating TIV from bland thrombus. A cross-validated multi-feature model did not improve diagnostic performance.