PET Image Radiomics Feature Variability in Lung Cancer: Impact of Image Segmentation

Radiomic features are powerful tools for characterizing intra-tumor heterogeneity and prognostic modeling. Yet, the lack of robustness evaluation poses difficulties in their application across institutions and among patient communities. The current study aimed to examine the impact of different image segmentation techniques on the extracted PET radiomics features in lung cancer. To this end, 120 patients with non-small cell lung cancer were enrolled. Different segmentations were applied to the images, including iterative thresholding using various threshold (30%, 40%, 60% and 80%), Fuzzy C-means (FCM), region growing, and manual delineation. For each of the contours, 841 features belonging to shape and first-, second-, and higher-order statistics features were extracted. The intraclass correlation coefficient (ICC) was calculated to assess features variability and selecting robust features. The results demonstrated that first-order features showed top performance in terms of robustness against various segmentations. In contrast, Gray Level Co-occurrence Matrix (GLCM), Gray Level Dependence Matrix (GLDM), Gray Level Size Zone Matrix (GLSZM), and shape-based parameters achieved poor performance. In addition, wavelate-based LLL set features achieved the highest robustness. The performance of radiomics features in terms of robustness depends on radiomics features set and features family. Hence, the more robust features should be considered in prognostic modeling.