Lung Nodules Detection System Using Support Vector Machine Classifier Combined Linear Discriminant Analysis-Based Feature Selection with Rule-Based Feature Pruning

Early nodule detection is significant for the diagnosis and clinical treatment of lung cancer. An efficient computer-aided detection system is developed to detect lung nodules in computed tomography scan image. In order to highlight lesion area, lung parenchyma segmentation including bronchial removal and contour pruning is implemented by iterative threshold and rolling ball algorithm. Mean-shift algorithm is applied to further smooth and enhance inner-structures of lung parenchyma. To effectively reduce false positive nodules, hybrid features are extracted using the rule-based feature pruning technology, they are regarded as input samples of SVM classifier to distinguish nodules from nodule candidates. Numerous experiments are conducted on a large dataset from lung image database consortium by various classifies, the diagnosis results demonstrate that the proposed nodule detection system achieves a promising classification accuracy, sensitivity and specificity in overall performance.