Detection of Lung Nodules in CT Scans Based on Unsupervised Feature Learning and Fuzzy Inference

Timely diagnosis of lung cancer is vital because of its increasing rate of occurrence and because of its high mortality rate. The large number of slices in CT scans and the different features of the nodules themselves make detection time-consuming for radiologists. The present study implemented an automated method for detection of lung nodules in CT scans based on an unsupervised feature learning technique. CT scans from 22 patients were used to evaluate the results of the method. The data was provided by the LIDC database (Lung Image Database Consortium) and the number of nodules was different in different images. The proposed approach comprised five steps. First, threshold measurement was used in an unsupervised manner to separate the lungs from other tissues. Next, nodules in the lungs were detected by their spherical shape and by analyzing the eigenvalues of the Hessian matrix. Third, a fuzzy inference system was used to learn features and to decrease system error. Fourth, candidate regions were classified and fifth, by using a hybrid system (combined the results of several classifiers), the error was further decreased. Shape features were used to train the classifier to detect lung nodules. The sensitivity of the proposed system was 85.5% and the accuracy of the proposed algorithm was 99.7% with a false positive rate of 1.99. The proposed algorithm was tested on 41 nodules and was reasonably accurate.