Classification of Lung Nodules Using Textural Features

Abstract

In this study, a computer-aided detection system was developed for detection and investigation of lung nodules in computed tomography images using image processing techniques. The computer aided detection system consists of four stages. First and second stages are preprocessing stages. First preprocessing stage is two-dimensional preprocessing stage and second preprocessing stage is three-dimensional preprocessing stage. Third stage of the developed system is the feature extraction stage. Five different groups of features are extracted from volume of interests in this stage. Last stage is the the nodule detection stage. The support vector machine algorithm is optimized using evolutionary algorithms to classify volume of interests using the features. The computer aided detection system achieves 93.76% sensitivity, 81.69% selectivity, 84.52% accuracy and 3.63 false positive per scan using only morphologic, statistical and histogram based features. After the inclusion of groups of outer surface statistical and outer surface GLCM and Gabor filter based textural features, performance rates of the computer aided detection system reaches 98.35% sensitivity, 90.42% selectivity, 92.28% accuracy and 2.33 false positive per scan. Results of experiments reveal that outer surface statistical and textural features are useful to increase sensitivity of the system. These features also decrease the number of false positives of the developed systemi