Higher-order CRF tumor segmentation with discriminant manifold potentials

The delineation of tumor boundaries in medical images is an essential task for the early detection, diagnosis and follow-up of cancer. However accurate segmentation remains challenging due to presence of noise, inhomogeneity and high appearance variability of malignant tissue. In this paper, we propose an automatic segmentation approach using fully-connected higher-order conditional random fields (HOCRF) where potentials are computed within a discriminant Grassmannian manifold. First, the framework learns within-class and between-class similarity distributions from a training set of images to discover the optimal manifold discrimination between normal and pathological tissues. Second, the conditional optimization scheme computes non-local pairwise as well as pattern-based higher-order potentials from the manifold subspace to recognize regions with similar labelings and incorporate global consistency in the inference process. Our HOCRF framework is applied in the context of metastatic liver tumor segmentation in CT images. Compared to state of the art methods, our method achieves better performance on a group of 30 liver tumors and can deal with highly pathological cases.