Performance Analysis of Feature Selection Techniques for Support Vector Machine and its Application for Lung Nodule Detection

Lung cancer typically exhibits its presence with the formation of pulmonary nodules. Computer Aided Detection (CAD) of such nodules in CT scans would be of valuable help in lung cancer screening. Typical CAD system is comprised of a candidate detector and a feature-based classifier. In this research, we study and explore the performance of Support Vector Machine (SVM) based on a large set of features. We study the performance of SVM as a function of the number of features. Our results indicate that SVM is more robust and computationally faster with a large set of features and less prone to over-training when compared to traditional classifiers. In addition, we also present a computationally efficient approach for selecting features for SVM. Results are presented for a publicly available Lung Nodule Analysis 2016 dataset. Our results based on 10-fold validation indicate that SVM based classification method outperforms the fisher linear discriminant classifier by 14.8%.