Feature Extraction and Analysis for Lung Nodule Classification using Random Forest

Early detection of lung nodule decreases the risk of advanced stages in lung cancer disease. Random forest (RF), a machine learning classifier, is used to detect the lung nodules and classify soft-tissues into nodules and non-nodules. A lung nodule classification approach is proposed to improve early detection for nodules. A five stages model has been built and tested using 165 cases from the LIDC database. Stage 1 is image acquisition and preprocessing. Stage 2 is extracting 119 features from the CT image. Stage 3 is refining feature vectors by removing all duplicate instances and undersampling the non-nodule class. Stage 4 is tuning the RF parameters. Stage 5 is examining different collections from the extracted feature sets to select those scores best for classification. The accuracy achieved by RF is the highest compared to other machine learning classifiers such as KNN, SVM, and DT. The proposed method aimed to analyze and select features that maximize classification results. Pixel based feature set and wavelet-based set scored best for higher accuracy. RF was tuned with 170 trees and 0.007 for in-bag fraction. Best results were achieved by the proposed model are 90.67%, 90.8% and 90.73% for sensitivity, specificity, and accuracy respectively.