Machine learning quantification of hepatobiliary lesion-to-liver contrast enhancement ratio in benign hepatocellular
Édouard Reizine1,2,3, Salim Jazzar1, Sonia Colombo4
1Department of Radiology, APHP, centre hospitalier universitaire Henri-Mondor, Créteil, France.
Objectives:
To evaluate the feasibility of a machine learning derived semi-automatic quantification of the lesion-to-liver contrast enhancement ratio following hepatobiliary contrast agent injection for the characterization of benign hepatocellular tumors.
Materials And Methods:
Precontrast, arterial, portal venous and hepatobiliary phases of liver magnetic resonance images were coregistered, then the liver was automatically segmented by a pretrained U-Net network. Each lesion was then manually segmented on the arterial phase, while different regions of interest were automatically placed in the liver parenchyma (ring shape, circular). Two-dimensional and three-dimensional lesion-to-liver contrast enhancement ratios were automatically calculated using each different regions of interest and the whole liver volume. All the calculated lesion-to-liver contrast enhancement ratio were compared to a manual lesion-to-liver contrast enhancement ratio and correlated to organic-anion-transporting polypeptide expression. Moreover, the ability of each calculated lesion-to-liver contrast enhancement ratio in classifying the two lesions was investigated.
Results:
A total of 23 hepatocellular benign tumors were analyzed including nine focal nodular hyperplasias (39 %) and 14 hepatocellular adenomas (61 %). The correlation between each calculated lesion-to-liver contrast enhancement ratio and the manual value was strong (Pearson coefficient ranging between 0.69 and 0.87). Regarding correlation with organic-anion-transporting polypeptide expression, the highest correlation values was obtained with two-dimensional ring region of interest lesion-to-liver contrast enhancement ratio (Pearson coefficient = 0.56). All lesion-to-liver contrast enhancement ratio values were effective in classifying between focal nodular hyperplasia and hepatocellular adenoma, with the best results in terms of inverse Fisher discriminant ratio obtained with ring-shaped region of interest (0.34 2D, 0.28 3D) and uniform liver region of interest (0.28 2D, 0.37 3D).
Conclusion:
Semi-automatic quantification of the hepatobiliary lesion-to-liver contrast enhancement ratio both using a two- or a three-dimensional approach is a promising tool for benign hepatocellular characterization.
