Skip to Main Content
In general, an automatic lung nodule detection system consists of two stages: (1) detection of objects within lung images as potential nodules (2) classification of the detected objects into nodule and nonnodule classes. This paper addresses the first stage by introducing a new method for shape based segmentation of 3D lung images. Firstly, the 3D geometric features of each voxel are calculated by using the partial derivatives of the 3D image, e.g. the Gaussian and mean curvature, principal curvatures, and shape index; Secondly, the shape features of the isointensity surfaces are subsequently extracted; Finally, a hybrid methodology incorporating shape feature extraction and 3D intensity-based region growing is applied to give accurate separation of connected objects having different shapes but similar intensity values. The experimental results from six CT scans demonstrate that the proposed method yields a high performance of nodule detection, (30 nodules out of 33 were correctly detected, a detection rate of about 91%), with reasonable false positive (FP) (average FP is about 1.29/slice), which can be further reduced by the classification stage. Moreover, unlike the traditional intensity-based method, using the proposed shape based method all of the nodules can be separated accurately from adjoining blood vessels or from the lung wall.