Shape-based ct lung nodule segmentation using five-dimensional mean shift clustering and MEM with shape information

This paper presents a joint spatial-intensity-shape (JSIS) feature-based method for the segmentation of CT lung nodules. First, a volumetric shape index (SI) feature based on the second-order partial derivatives of the CT image is calculated. Next, the SI feature is combined with spatial and intensity features to form a five-dimensional feature vectors, which are then clustered using mean shift to produce intensity and shape mode maps. Finally, a modified expectation-maximization (MEM) algorithm is applied on the mean shift intensity mode map to merge the neighboring modes with spatial and shape mode maps as priors. The proposed method has been evaluated on a clinical dataset of thoracic CT scans that contains 80 nodules. A volume overlap ratio between each segmented nodule and the ground truth annotation is calculated. Using the proposed method, the mean overlap ratio over all the nodules is 0.81 with standard deviation of 0.05. Most of the nodules, including challenging juxta-vascular and juxta-pleural nodules, can be properly separated from adjoining tissues.