Lung Cancer Recurrence Prediction Using Radiomics Features of PET Tumor Sub-Volumes and Multi-Machine Learning Algorithms

Abstract

We aimed to predict recurrence in lung cancer patients using PET radiomics features and machine learning algorithms. In this work, 136 non-small cell lung cancer (NSCLC) patients were enrolled. To study the impact of tumor sub-volume on recurrence prediction's accuracy, five sub-regions or contours were delineated manually, were extended with different distances (1, 2, 3, 4, and 5 mm). Three different feature selections and ten classifiers with 100 bootstraps were utilized. Our results illustrated that contourPlus1mm with the Minimum Redundancy Maximum Relevance (mrmr) feature selection and Random Forest (RF) classifier, contourPlus1mm with the MRMR feature selection and Linear Discriminant Analysis (LDA) classifier, and contourPlus4mm with the Recursive Feature Elimination (RFE) feature selection and Logistic regression (LR) classifier, had the highest performance (AUC= 0.65). The results of this study illustrated that an extended sub-volume of a manual contour boosts the performance of recurrence prediction in patients with lung cancer. This study demonstrated that the use of contour extending method can be effective in increasing the predictive accuracy of different machine learning classifier methods.