4.3. Tumor Delineation, Image Pre-Processing and Extraction of Radiomics Features

Using Osirix Lite (Pixmeo SARL, Bernex, Switzerland) and MiM software (version 7.0.4, MIM Software Inc., Cleveland, OH, USA), regions of interests (enhancing tumor on T1+Gadolinium images and combined tumor/edema portion on T2-weighted images) were manually delineated on all diagnostic MRI-images of the training and validation cohort, supervised by two experienced neuro-radiation oncologists (DE, IC).

Using Python 3.7 and the dedicated packages (cv2 version 4.1.0, https://pypi.org/project/opencv-python/, (accessed on 23 December 2020)), SimpleITK version 1.2.0 (https://simpleitk.org/, (accessed on 23 December 2020)) and scikit-image version 0.14.2, (https://scikit-image.org/, (accessed on 23 December 2020)), an image pre-processing routine was developed to handle the broad variability of image acquisition and reconstruction parameters.

At first, spatial resolution of the images was normalized with respect to the image sequence (final pixels are: 0.449 mm² and slice thickness of: 5.5 mm). The mode of the pixel spacing and slice thickness distributions from the Maastricht UMC+ cohort were used as reference values for the resampling procedure to minimize the number of resampled images. A bicubic interpolation over 4 × 4 pixel neighborhood was used for both upsampling and downsampling. In order to correct the low frequency intensity non-uniformity, which is intrinsic for MRI images, the N4 bias field correction algorithm was used [68]. Furthermore, the histogram equalization method implemented in the scikit-image 0.15.0 package [69] was used to enhance the contrast of MRI images [70]. As the last step of the pre-processing routine, image intensities were normalized using Z-score standardization method [71]. A pre-processing routine was applied to both cohorts, where parameters (mu, sigma) for the Z-score transformation were evaluated on the training cohort and transferred to the validation cohort. Parameters used are T1 mu = 0.1904, T1 sigma = 0.2313, T2 mu = 0.2009 and T2 sigma = 0.2448.

In order to obtain the quantitative imaging features, an open-source Pyradiomics 2.2.0 python package for the radiomics features extraction was utilized [72]. Using the dedicated MRI settings, features from following feature classes were extracted: First Order Statistics, Shape-based (2D and 3D), Gray Level Cooccurence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Gray Level Dependence Matrix (GLDM), Neighboring Gray Tone Difference Matrix (NGTDM). Along with the original features Laplacian of Gaussian (LoG) (sigma: (2.0,3.0,4.0,5.0) and Wavelet filters were activated resulting in a total of 1197 features per patient. A detailed mathematical feature description as provided by Aerts et al. 2014 [30].