Tumor detection in CT liver images is a challenging task. The nature of tumor has a direct effect on the number of voxels being contaminated, as well as on the changes in the observed CT scan. In order to deal with this challenge, in this paper we propose the use of advanced non-linear machine learning techniques to determine the optimal features, as well as the hyperplane that use these features to separate tumoral voxels from voxels corresponding to healthy tissues. Very promising classification results using an important volume of clinically annotated data (86% sensitivity, 82% specificity) demonstrate the potentials of our approach.