Milea Timbergen

145 Supplemental Material 1. Radiomics feature extraction This Supplemental material is similar to 1 , but details relevant for the current study are highlighted. A total of 411 radiomics features were used in this study. All features were extracted using the defaults for MRI scans from the Workflow for Optimal Radiomics Classification (WORC) toolbox 2 , which internally uses the PREDICT 3 and PyRadiomics 4 feature extraction toolboxes. The code to extract the features for this specific study has been published open- source 5 . An overview of all features is depicted in Supplemental Table 2. For details on the mathematical formulation of the features, we refer the reader to Zwanenburg et al. (2020) 6 . More details on the extracted features can be found in the documentation of the respective toolboxes, mainly the WORC documentation 7 . The features can be divided in several groups. Twelve histogram features were extracted using the histogram of all intensity values within the Regions of Interest (ROIs), i.e. the tumours, and included several first-order statistics such as the mean, standard deviation and kurtosis. To create the histogram, the images were binned using a fixed number of 50 bins. Seventeen shape features were extracted based only on the ROI, i.e. not using the image, and included shape descriptions such as the volume, compactness, roundness and circular variance. The orientation of the ROI was described by three features, which represent the three major axis angles of a 3-D ellipse fitted to the ROI. Lastly, 379 texture features were extracted using the Gray Level Co-occurrence Matrix (144 features), Gray Level Size Zone Matrix (16 features), Gray Level Run Length Matrix (16 features), Gabor filters (72 features), Laplacian of Gaussian filters (36 features), vessel (i.e. tubular structure) filters (36 features) 8 , local phase filters (36 features) 9 , Local Binary Patterns (18 features), and the Neighborhood Grey Tone Difference Matrix (5 features). Most of the texture features include parameters to be set for the extraction. Beforehand the values of the parameters which will result in features with the highest discriminative power for the classification at hand (e.g. DTF vs non-DTF) is not known. Including these parameters in the workflow optimization, see Supplemental Material 2, would lead to repeated computation of the features, resulting in a redundant decrease in computation time. Therefore, alternatively, these features are extracted at a range of parameters as is default in WORC. The hypothesis is that the features with high discriminative power will be selected by the feature selection methods and/or the machine learning methods as described in Supplemental Material 2. The parameters used are described in Supplemental Table 1. 5