This paper proposes a system to extract hepatocellular carcinomas from two-phase three-dimensional abdominal CT images, or early phase image and late phase image. The process consists of four steps ; 1)extraction of liver region, 2)enhancement of cancer regions, 3)extraction of candidates, and 4)feature measurement and classification of the candidates whether it corresponds to cancer or not. The salient feature of this system is that the processes make the most of information from two phase images in order to detect cancers accurately. We have applied the system to three dimensional abdominal CT images from fifteen patients obtained with multi-detector row CT scanners and confirmed that false positives per case estimated by leave-one-out method is 0.53 for Support Vector Machine and 0.13 for Mahalanobis distance based classifier when the sensitivity is 100 %.