Metastatic Liver Tumor Segmentation Using Texture-Based Omni-Directional Deformable Surface Models

The delineation of tumor boundaries is an essential task for the diagnosis and follow-up of liver cancer. However accurate segmentation remains challenging due to tissue inhomogeneity and high variability in tumor appearance. In this paper, we propose a semi-automatic liver tumor segmentation method that combines a deformable model with a machine learning mechanism. More precisely, segmentation is performed by an MRF-based omni-directional deformable surface model that uses image information together with a two-class (tumor, non-tumor) voxel classification map. The classification map is produced by a kernel SVM classifier trained on texture features, as well as intensity mean and variance. The segmentation method is validated on a metastatic tumor dataset consisting of 27 tumors across a set of abdominal CT images, using leave-one-out validation. Compared to pure voxel and gradient approaches, our method achieves better performance in terms of mean distance and Dice scores on the group of 27 liver tumors and can deal with highly pathological cases.