Kidney segmentation in 3D CT images using B-Spline Explicit Active Surfaces

Abstract

In this manuscript, we propose to adapt the B-Spline Explicit Active Surfaces (BEAS) framework for semi-automatic kidney segmentation in computed tomography (CT) images. To study the best energy functional for kidney CT extraction, three different localized region-based energies were implemented within the BEAS framework, namely localized Chan-Vese, localized Yezzi, and signed localized Yezzi energies. Moreover, a novel gradient-based regularization term is proposed. The method was applied on 18 kidneys from 9 CT datasets, with different image properties. Several energy combinations were contrasted using surface-based comparison against ground truth meshes, assessing their accuracy and robustness against surface initialization. Overall, the hybrid energy functional combining the localized signed Yezzi energy with gradient-based regularization simultaneously showed the highest accuracy and the lowest sensitivity to the initialization. Volumetric analysis demonstrated the feasibility of the method from a clinical point of view, with similar reproducibility to manual observers.