KIDNEY SEGMENTATION FROM CT IMAGES USING A 3D NMF-GUIDED ACTIVE CONTOUR MODEL

A hybrid framework that is based on the integration of geometric deformable models and nonnegative matrix factorization (NMF) is introduced for 3D kidney segmentation from abdominal computed tomography (CT) images. The NMF is employed due to its ability to cluster complex data by extracting discriminative features from higher dimensional space. In this paper, regional features from CT appearance, a kidney shape model, and spatial interactions are fused using the NMF to produce a more robust model to guide the deformable model's evolution. The shape model is constructed using a set of training images and is updated during segmentation using an appearance-based method taking into account both voxels' locations and appearances. The spatial interactions are modeled using a pair-wise Potts Markov-Gibbs random field model. Our approach has been tested on 36 in-vivo 3D CT data sets and evaluated using, the Dice coefficient, the 95-percentile Hausdorff distance, and percentage kidney volume difference. Evaluation results show that feature fusion using NMF increases the deformable model's ability to accurately segment complex CT kidney data.