An Automated Initialization System for Robust Model-Based Segmentation of Lungs in CT Data

Abstract. In this paper we study the feasibility of an automated ini-tialization system for a robust model-based lung segmentation approach. The lung segmentation method consists of a robust active shape model (RASM) matching stage followed by an optimal surface finding step. The RASM needs to be initialized in rough proximity to the target structure for achieving good segmentation results. For this purpose we employ a feature-based alignment (FBA) method to predict three landmark points (lung apex and carina). These landmark points are subsequently used to calculate initial model parameters. The approach was evaluated on a di-verse set of 98 CT scans of normal and diseased lungs by comparison to an independent reference standard. For segmentation results based on the new initialization approach, the mean and standard deviation of the Dice coefficient was 0.976 ± 0.025. In comparison, the Dice coeffi-cient of lung segmentations based on the previously utilized initialization method was 0.971 ± 0.040. The new RASM initialization approach is 3-times faster than the previously utilized method and can be expanded by including additional landmarks