Radiomics Feature Extraction from Lung Phantom

An advanced imaging software package (Mirada RTx 1.6, Mirada Medical, Oxford, UK) was used for importing, exporting and contouring purposes. The shredded rubber and ABS20 cartridges of the CCR phantom were predominantly used. The rubber cartridge was chosen because it was reported to have HU values characteristics similar to non-small cell lung cancer (NSCLC) tumors.^{15 (link)} An automatic contouring tool in Mirada RTx was used to contour ROIs. A spherical ROI of volume 4.2 cm³ was contoured on the central region of each cartridge and kept identical across all scanners. Radiomics features were extracted using an in-house program. The features were composed of shape descriptors (10), intensity histogram statistics (16), gray level co-occurrences matrices (GLCM, 24), grey level run length matrices (GLRLM, 11), grey level size zone matrix (GLSZM, 11), neighborhood grey tone difference matrix (NGTDM, 5), fractal dimensions (8) and intensity histogram wavelets (128) for a total of 213 features. Intensity volume histograms were used to calculate the first order features. Second order features based on GLCM were initially developed by Haralick et al.^{24 , 25} These features were implemented in our program as described by Aborisade et al.²⁶ Volumetric interpretation of texture features were given by Arati et al.²⁷ GLCM features provide spatial dependence of neighboring voxels as described by Oliver et al.^{28 (link)} GLRLM features were implemented according to definitions provided by Galloway, Chu et al., and Dasarathy and Holder.^{29 –31} GLSZM and NGTDM based features were first developed by Thibault et al., and Amadasun et al. respectively.^{32 , 33} Fractal dimensions features were calculated as described by Sarkar et al., and Jin et al.,^{34 , 35} A biorthogonal basis function was applied to the original and resampled CT images. A combination of a one-dimensional low pass and a high pass filters applied to a three dimensional image generated 8 wavelet filtered data sets. The first order wavelet features were then extracted from these data sets as described by Aerts et. al.³ Sixty four equispaced gray levels (Ng = 64) were used to discretize the intensities of image voxels for calculating all features unless otherwise specified.