Topology constraint graph-based model for non-small-cell lung tumor segmentation from PET volumes

We propose a graph-based segmentation model that takes into account both topological information and intensity similarity to segment non-small-cell lung carcinoma (NSCLC) from PET volumes. The proposed model estimates the probabilities of each voxel belonging to the given foreground and background labels. The topological information is derived from our region of interest topological skeleton tree model (ROI-TKT) to better capture the inclusion and nesting information. Topological information contributes to the separation of tumors from adjacent lesions with similar intensities and the delineation of tumors with heterogeneous density distributions. The experimental evaluation on 20 PET volumes with manual tumor segmentation results from two doctors demonstrated that our method outperformed two other graph-based segmentation methods as measured by the Dice similarity coefficient (DSC).