Lung Nodule Detection Using Polygon Approximation and Hybrid Features from CT Images

Background: Lung cancer is one of the leading causes of cancer-related deaths. Computer-aided diagnosis systems facilitate the radiologists in early diagnosis of pulmonary nodules. It improves the performance of the diagnostic process by providing the second opinion. Methods: A lung nodule detection method which consists of multiple steps has been proposed. At first step lung region is separated using the optimum multi-thresholding method. For smoothing the boundaries and filling the holes, morphological operations are used. In the next step, nodule candidates are extracted using polygon approximation. Further, features of nodule candidates are extracted, and a hybrid feature vector is created using Histograms of Oriented Gradients, intensity and geometric features from enhanced nodule candidates, supported by Contrast Limited Adaptive Histogram Equalization. On the basis of selected feature vectors, Support Vector Machine has been used for classification. Conclusion: The proposed system is evaluated over a standard dataset of Lung Image Consortium Database (LIDC) and achieved an accuracy of 98.8%, sensitivity 97.7%, specificity 96.2% and very low false positive rate of 3.8. It will provide valuable assistance to the radiologists in the diagnostic process.