Stability Analysis of CT Radiomics Features With Respect to the Variation of Manual Segmentation in Oropharyngeal Cancer

Abstract

Accurate segmentation of tumors and quantification of tumor features are important for cancer detection, diagnosis, monitoring, and planning therapeutic intervention. Due to inherent noise components in multiparametric imaging and inter-observer variations, it is common that various segmentation methods, including both manual segmentation and computer-aided segmentation methods, may produce large segmentation errors in tumor volumes and their associated radiomics features. Such errors may eventually lead to large prediction errors. The aim of this study is to carry out the stability analysis for various radiomics features with respect to segmentation results based on the contrast-enhanced CT axial images. All the 436 CT images were imported to a commercial software and the gross tumor volumes of primary disease (GTVp) were segmented by two expert radiation oncologists. In order to illustrate the variations of segmentation results, additional two segmentation results were set up via resizing the original segmented regions of interest (ROIs) based on their geometric information on the surface. For the three ROI image groups, we calculated 109 radiomics features, and wavelet-based features. Then, a logistic regression model was built to investigate the correlation between the radiomics features extracted from GTVp and the response to chemoradiation in terms of overall survival (OS). Finally, based on the prediction probabilities, we assessed the inter-rater reliability and reproducibility via calculating the intraclass correlation coefficients (ICC) and concordance correlation coefficients (CCC), respectively. One of the features, 25percentile, outperforms other features in terms of both ICC and CCC. Its ICC and CCC between the original segmentation group and resized segmentation groups are around 0.27. In terms of the prediction accuracy, top features with high AUC extracted from different segmentation results are different from each other. Another interesting finding is that gray-level co-occurrence matrix (GLCM) based features have low ICC and CCC (<0.3) in comparison between original segmentation group and each resized segmentation group, while have high ICC and CCC (>0.9) in comparison between two resized segmentation groups. Considering the prediction performance, these GLCM based features derived from original segmentation group perform well with the area under the curve (AUC) above 0.7, which means these traditional features are sensitive to segmentation results. The segmentation variability affects both the radiomics features and prediction accuracy. Moreover, it is of great meaning to discover radiomics features with robustness of segmentation variability in oropharyngeal cancer, which can be warranted for treatment monitoring and prognosis prediction.