Lung nodule detection based on GA and SVM

To solve the problem that the Juxta-Vascular nodule and vascular-crossing could not be distinguished by dot filter and make high rate of false positive, a feature subset selection method based on improved genetic algorithms in wrapper model was proposed, and the best feature subset was used to establish a classifier based on support vector machines to improve the performance by reducing false positive and retaining true nodule. From 22 features (including three newly proposed features) which were calculated for each detected structure, seven features were selected as the optimal feature subset. And the classifier trained with the optimal feature subset resulted in 100% sensitivity and 95.5% specificity from the lung nodule database (50 true nodules and 961 false ones). Experiments show that the framework and the approach can be applied in clinic and ease the workload of the radiologist in interpreting lung CT scans; at the same time, the same method can also be applied to other computer-aided detection fields such as detection of mammary tumor.