Segmentation agreement and the reliability of radiomics features

Radiomics features extracted from medical images have been shown to correlate with tumor histological biomarkers and patient clinical information. An accurate selection of reliable features is crucial for an efficient prediction/stratification analysis in clinical applications. This work introduces a computational method for evaluating the reliability of radiomics features with respect to image segmentation. To do so, we define four reliability scores that link segmentation variability to feature quality, consistency, robustness and instability. These scores can be used to establish a ranking that helps identifying the best features to employ for the application at hand, allowing an a-priori evaluation of radiomics reliability in clinical applications. We show the performance of the method with Magnetic Resonance images of meningioma patients. In this case, we identified three main groups of features displaying similar behaviors with respect to image segmentation variability.