The domain knowledge based graph-cut model for liver CT segmentation

In this paper, we present a semi-supervised approach for liver segmentation from computed tomography (CT) scans, which is based on the graph cut model integrated with domain knowledge. Firstly, some hard constraints are obtained according to the knowledge of liver characteristic appearance and anatomical location. Secondly, the energy function is constructed via knowledge based similarity measure. A path-based spatial connectivity measure is applied for robust regional properties. Finally, the image is interpreted as a graph, afterwards the segmentation problem is casted as an optimal cut on it, which can be computed through the existing max-flow algorithm. The model is evaluated on MICCAI 2007 liver segmentation challenge datasets and some other CT volumes from the hospital. The experimental results show its effectiveness and efficiency.

► The graph cut framework includes both regional and boundary properties of a segment. ► Domain knowledge can be exploited to constraint the solution space of segmentation. ► The path-based spatial connectivity measure brings about quality regional properties. ► The proposed model efficiently achieves robust global optima of the energy function.