Lymphovascular Invasion Prediction in Lung Cancer Using Multi-Segmentation PET Radiomics and Multi-Machine Learning Algorithms

Abstract

The current study aimed to predict the lymphovascular invasion in lung cancer patients using PET radiomics features and multivariate analysis of various segmentation techniques. To this end, 126 patients with non-small cell lung cancer (NSCLC) that underwent PET/CT examinations were enrolled in this study protocol. Multiple segmentations were applied to the images, including k-Means, iterative thresholding with different thresholds, local active contour, watershed, and manual contouring of lesions. For each dataset, a total 105 features belonging to shape and first-order statistics and texture features were extracted. Three feature selection and ten classifiers with 100 bootstraps were utilized through various segmentation methods using machine learning algorithms. The results demonstrated that Local Active Contour (LAC) methods with recursive Feature Elimination (RFE) feature selection and Naive Bayes (NB) classifier had the highest performance (AUC = 0.94), followed by K-means method with the Minimum Redundancy Maximum Relevance (MRMR) feature selection and NB classifier (AUC = 0.93). Different segmentation algorithms achieve different performance. The radiomics features extracted from ROIs of various segmentation algorithms are suitable biomarkers for predicting the lymphovascular invasion in patients with NSCLC.