Automated detection and classification of pulmonary nodules in 3D thoracic CT images

Abstract

Computer-aided diagnosis (CAD) systems for recognition of pulmonary nodules are of great importance for early diagnosis of lung cancer. In this paper, we propose a CAD system for automated detection and classification of pulmonary nodules in 3D computerized-tomography (CT) images. First, we segment lung parenchyma from the CT data using a thresholding method. Afterwards, we apply Gaussian filters for noise reduction and nodule enhancement. Then, we use intensity and volumetric shape Index (SI) for detecting suspicious nodule candidates that include both nodules and vessels. Besides, by using SI, we can recognize nodules that are attached to vessels, pleural wall or mediastinal surface. Next, features such as sphericity, mean and variance of the gray level, elongation and border variation of potential nodules are extracted to classify detected nodules to malignant and benign groups. Fuzzy KNN is employed to classify potential nodules as non-nodule or nodule with different degree of malignancy. To assess our proposed method, 63 thoracic CT scans, from the Lung Image Database Consortium (LIDC), are recruited. Our method achieved sensitivity of 88% for nodule detection with approximately 10.3 False-Positive (FP)/subject; also the achieved classification of nodules is concordant with radiologists' opinion. Considering nodules of small size, as well as those with irregular shape, the results of nodule detection and classification are reasonable.