Differentiation Between Mass-forming Type Peripheral Cholangiocarcinoma and Hepatic Abscesses: Application of Artificial Neural Networks to CT Images

ABSTRACT

PURPOSE: To determine which CT findings are useful for differentiating cholangiocarcinomas (CC) from hepatic abscesses and also to determine whether artificial neural networks (ANNs) improve radiologists' performance. MATERIALS AND METHODS: CT findings of 51 patients with mass-forming type CC and 70 patients with hepatic abscesses were analyzed with morphologic, enhancing and other ancillary findings by three radiologists with differing levels of expertise independently. ANNs were constructed using statistically significant CT findings derived from the analyses. The performances of the ANNs and the radiologists were evaluated using receiver operating characteristic analysis. RESULTS: CT findings of rim-like enhancement, lymphadenopathy, capsular retraction, focal bile duct dilatation and a solid component were significant features of CC (p< 0.05). Findings of a clustered sign, multilayered enhancement, sharp margin, round shape, and air-biliary gram were significant features of hepatic abscesses. The ANNs showed better performance (AZ=0.9673, 98.0%, 97.1%, and 97.5%, respectively) than the resident (AZ=0.898, 78.4%, 81.4%, 80.2%) (p<0.05) in differentiating between the two diseases (AZ, sensitivities, specificities, and overall accuracies). However, there were no significant differences in the diagnostic performance of the ANNs and the two board-certified radiologists. CONCLUSION: Several CT findings are useful in differentiating CC from hepatic abscesses and ANNs may improve the performance of a radiologist with little experience.