A soft-computing based approach towards automatic detection of pulmonary nodule

Abstract

Early detection of lung cancer is the major challenge for physicians to treat and control this deadly disease whose primary step is to detect pulmonary nodule from thoracic computed tomography (CT) images. In view of increasing the accuracy of the pulmonary nodule detection methodology, this paper proposes a novel technique that can aid early diagnosis of the patients. The study has considered high resolution computed tomography (HRCT) images from two public datasets LIDC and Lung-TIME and an independent dataset, created in collaboration between Peerless Hospital Kolkata and University of Calcutta. The key feature of the test dataset is that the class features are imbalanced in nature. The structures associated with lung parenchyma are segmented using parameterized multi-level thresholding technique, grayscale morphology, and rolling ball algorithm. Then random under sampling is implemented to overcome the imbalance class problem, followed by a feature selection methodology using binary particle swarm optimization (BPSO). The nodule and non-nodule classification are performed by implementing ensemble stacking. Indeed, it has been observed that there exists insufficient published literature that has been considered similar looking pulmonary abnormalities as non-nodule objects as well as imbalance class problem and feature selection algorithm to design an automated, accurate and robust model for automated detection of the pulmonary nodule. In reference to the LIDC dataset, the false positive, false negative detection rates and sensitivity are 1.01/scan, 0.56/scan and 99.01% respectively, which is an improvement in terms of accuracy as compared to the existing state-of-the-art research works.