| In this work, several separate but related topics in computer-aided detection (CAD) of pulmonary nodules in helical CT scans were investigated. These included curvature computation of level surfaces in 3D voxel images, a hybrid CAD system of locating juxtaplerual and non-pleural nodules individually, a nodule candidate generation CAD scheme which deals with juxtaplerual and non-pleural nodules simultaneously.; Curvature formulas which are symmetric with respect to coordinates were developed. By choosing the normalized gradient of the level function as the reference unit normal of the level surface, unique and consistent geometrical meaning of the curvature sign is maintained. Furthermore, curvature computation formulas are expressed in concise matrix forms that are easier to handle.; A hybrid CAD system for finding small-sized pulmonary nodules in high resolution helical CT scans was developed. For juxtapleural nodules, morphological closing, thresholding and 3D component analysis were used to obtain nodule candidates; gray level and geometric features were analyzed using linear discriminant analysis (LDA) classifier. To deal with non-pleural nodules, a discrete-time cellular neural network (DTCNN) based on local shape features was developed. To tailor it for lung nodule detection, this DTCNN was trained using genetic algorithms (GAs) to derive the shape index variation pattern of nodules.; Additionally, a new CAD system is proposed to locate both juxtapleural nodules and non-pleural nodules in a single procedure. Surface normal overlapping method is used to detect the roughly center position of the possible nodule candidate. This scheme was improved by combining the local shape property into the normal tracking procedure. |
