Robust versus Non-Robust Radiomic features: Machine Learning Based Models for NSCLC Lymphovascular Invasion

The application of radiomic features for predicting and diagnosing lung cancer has been established in previous studies. However, radiomic features might be weak in terms of reproducibility. This study aimed to determine robust features against lung motion in a multicenter study, followed by choosing bold features using various feature selection methods and comparing them with the standard method results without considering the robustness of features. To this end, an in-house developed lung phantom was used. A specific motor was manufactured to simulate motion in 2 orthogonal directions. Lesions and tumors were delineated using a semi-automatic segmentation. In total, 105 radiomic features were extracted. Three different feature selections and five machine learning classifiers were implanted in this study. First, the Intraclass Correlation Coefficient (ICC) was calculated to show the variability of radiomic features and select robust features with an ICC of more than 90%. Next, the selected robust features went through various feature selection methods. Finally, the results of this method were compared with outcomes of regular feature selection without considering the robust features in multiple machine learning classifiers of an imbalanced clinical data set. Our result demonstrated that although considering the robustness of radiomic features has a minor negative impact on prediction accuracy, it has a significant productive impact on sensitivity.