To evaluate the feasibility of a machine learning derived semi-automatic quantification of the liver-to-lesion contrast enhancement ratio (LLCER) following hepatobiliary contrast agent injection for the characterization of benign hepatocellular tumors.

. Pre-contrast, arterial, portal venous and hepatobiliary phases of liver MRI were co-registered, then the liver was automatically segmented by a pretrained UNET network. Each lesion was then manually segmented on the arterial phase, while different region-of-interests were automatically placed in the liver parenchyma (ring shape, circular). 2D and 3D LLCER were automatically calculated using each different ROI and the whole liver volume. All the calculated LLCER (cLLCER) were compared to a manual LLCER (mLLCER) and correlated to OATP expression. Moreover, the ability of each cLLCER in classifying the two lesions was investigated.

A total of 23 hepatocellular benign tumors were analyzed including 9 FNH (39%) and 14 HCA (61%). The correlation between each cLLCER and the mLLCER value was strong (Pearson coefficient ranging between 0.69 and 0.87). Regarding correlation with OATP expression, the highest correlation values was obtained with 2Dring ROI LLCER (Pearson coefficient = 0.56). All LLCER values were effective in classifying between FNH and HCA, with the best results in terms of Inverse Fischer Discriminant Ratio obtained with ring-shaped ROI (0.34 2D, 0.28 3D) and uniform liver ROI (0.28 2D, 0.37 3D).

Semi-automatic quantification of the hepatobiliary lesion to liver contrast enhancement ratio both using a 2D or a 3D approach is a promising tool for benign hepatocellular characterization.