Imaging and image analysis

The immunofluorescently stained microscope slides were fully digitized at 20x magnification using a digital slide scanner (Pannoramic Scan II, 3D HISTECH Ltd., Budapest, Hungary) equipped with a quad band (DAPI/FITC/TRITC/Cy5) filter set. DAPI filter was used for blue DAPI channel, FITC filter was used for green tissue autofluorescence channel, and TRITC filter was used for Ki67 channel. Images of the stained tissue slices were exported from slide scanner data sets (Pannoramic Viewer, Version 1.15.4, 3D HISTECH Ldt., Budapest, Hungary) as PNG images with pixel dimensions of 0.325 µm. Some regions in the exported images had to be masked by hand (Adobe Photoshop CS6, Adobe Systems Inc., San Jose, USA) in order to remove artifacts (i.e. tissue overlaps, air bubbles, unspecific staining, dirt/fluorescent particles, blooming, etc.). Spectral bleedthrough between different color channels was corrected using the “Spectral Unmixing” plugin for ImageJ (Version 1.51n, http://imagej.hih.gov/ij). Image analysis was performed with Mathematica (Version 11.1, Wolfram Research, Inc., Champaign, IL, USA). Corrected fluorescence images were imported and split into separate color channels. In order to obtain tissue masks (almost entirely represented by DAPI and autofluorescence signals), all images were smoothed with a 5 pixel wide Gaussian filter and binarized using Otsu’s (cluster variance maximization) thresholding method⁵⁵ prior to color channel separation. DAPI signals within blue image channels were also binarized using Otsu’s thresholding method while proliferation marker (Ki67) signals within red image channels were binarized using Kapur’s (histogram entropy minimization) thresholding method⁵⁶. Since specific proliferation marker staining can only occur within the nuclei, the binarized DAPI and Ki67 images were multiplied in order to omit unspecific staining outside of nuclei. The resulting masks were further cleared of very small segments (up to 20 pixels) to eliminate specks of fluorescent particles within nuclei. Finally, the areas of total tissue, DAPI and Ki67 masks were determined and ratios were computed. Numbers of analyzed images were as follows: 33 for untreated peritumoral brain tissue, 32 for peritumoral brain treated with TMZ + 4 Gy, 13 for untreated GBM tissue, 8 for GBM tissue treated with TMZ + 4 Gy.

To verify the result of the automated image analysis approach we performed an additional interactive analysis by three independent observers using ImageJ. Corrected fluorescence images were imported and split into separate color channels (DAPI, Ki67, autofluorescence). Subsequently, all color channels were segmented by interactive thresholding. Manually generated masks were imported in Mathematica and analyzed corresponding to the automatically segmented masks. Calculated parameters of the three observers’ segmentations were averaged and ratios were computed.

Tissue slices with apoptosis staining underwent the same imaging and image preprocessing procedures as the microscope slides stained against Ki67, as mentioned above. Apoptosis was captured using the TRITC filter of the digital slide scanner. Spectral unmixing was performed and apoptosis signals within red image channels were bianrized using Kapur’s (histogram entropy minimization) thresholding method. Binarized DAPI and apoptosis images were multiplied in order to omit unspecific staining outside of nuclei. Subsequently, segmented images were inspected and masked by hand if necessary (e.g. vessels, artifacts). Finally, the areas of total tissue, DAPI, and apoptosis masks were determined, ratios were computed, and results were averaged for all slices originating from the same tissue slice.