Abstract: In this paper, a novel kidney segmentation method for Computed Tomography patient data with kidney cancer is proposed. The segmentation process is based on Hybrid Level Set method with elliptical shape constraints. Using segmentation results, a fully automated technique of kidney region classification is introduced. Identification of the kidney, tumor and vascular tree is based on RUSBoost and the decision trees technique. This approach enables to resolve main problems connected with region classification: class imbalance and the number of voxels to classify. The classification is based on 64-element feature vectors calculated for the kidney region that consist of 3D edge region, orientation and spatial neighbourhood information. The proposed methodology was evaluated on clinical kidney cancer CT data set. Segmentation effectiveness in Dice coefficient meaning was equal to 0.85?0.04. Overall accuracy of the proposed classification model amount to 92.1% presented results confirm usefulness of the solution. We believe that this is the first solution which allow to segment (divide) kidney region into separable compartments, i.e. kidney, tumor and vascular trees.
Keywords: Segmentation, Computed Tomography, Neural Network, SVM, Decision Tree.