Automatic Segmentation of Pulmonary Nodules in CT Images

The accurate segmentation of pulmonary nodules lays the foundation for distinguishing malignant from benign pulmonary nodules. In this paper, a robust and automatic algorithm is proposed to segment lung nodules slice-by-slice in three dimensional (3D) Computed Tomography (CT) images. A nonparametric estimation method called Mean-Shift (MS) algorithm was applied to segmenting lung nodules. It is critical to set the proper bandwidth parameter in Mean-Shift algorithm. In this paper, a new bandwidth chosen method was presented. Imposing region-growing method and bandwidth selection theorem on extracting the initial smoothing bandwidth and multi-scale analyses and K-L divergence rule were used to determine the most proper bandwidth. And also, clustered using Mean-Shift in the nodule ROI defined by detection method was involved to get the accurate boundary. Moreover, this method was applied to clinical chest CT volumes containing 36 nodules (95 slices) and the proposed method segmented all of the nodules with only three false slices. Therefore, the approach presented can be consistently and robustly used to segment Ground Glass Opacity nodules, the nodules attached to lung walls and vessels and anisotropic nodules. The proposed method provides a powerful tool for automatic and accurate segmentation of nodules.