Deep Feature Selection and Decision Level Fusion for Lungs Nodule Classification

Deep Feature Selection and Decision Level Fusion for Lungs Nodule Classification

Abstract

The existence of pulmonary nodules exhibits the presence of lung cancer. The Computer-Aided Diagnostic (CAD) and classification of such nodules in CT images lead to improve the lung cancer screening. The classic CAD systems utilize nodule detector and feature-based classifier. In this work, we proposed a decision level fusion technique to improve the performance of the CAD system for lung nodule classification. First, we evaluated the performance of Support Vector Machine (SVM) and AdaBoostM2 algorithms based on the deep features from the state-of-the-art transferable architectures (such as; VGG-16, VGG-19, GoogLeNet, Inception-V3, ResNet-18, ResNet-50, ResNet-101 and InceptionResNet-V2). Then, we analyzed the performance of SVM and AdaBoostM2 classifier as a function of deep features. We also extracted the deep features by identifying the optimal layers which improved the performance of the classifiers. The classification accuracy is increased from 76.88% to 86.28% for ResNet-101 and 67.37% to 83.40% for GoogLeNet. Similarly, the error rate is also reduced significantly. Moreover, the results showed that SVM is more robust and efficient for deep features as compared to AdaBoostM2. The results are based on 4-fold cross-validation and are presented for publicly available LUNGx challenge dataset. We showed that the proposed technique outperforms as compared to state-of-the-art techniques and achieved accuracy score was 90.46±0.25%.