A total of 137 patients were included in this retrospective assessment of a prospective monocentric glioma registry. The inclusion criteria were the availability of a preoperative MRI and a detailed postoperative neuropathological diagnosis, including molecular diagnostics, which made a diagnosis according to the fifth WHO classification for CNS tumors 2021 possible [18]. All patients provided informed consent for the prospective glioma registry, which received approval from our local Institutional Review Board (IRB).

MRI scans were performed using a Philips 3 Tesla whole-body scanner (Achieva or Ingenia models, Philips, Best, The Netherlands). The Philips protocol includes an isotropic FLAIR (voxel size 1 mm3, Echo Time (TE) = 269 ms, Repetition Time (TR) = 4800 ms, Inversion Time (TI) = 1650 ms), isotropic T1w Turbo Field Echo (TFE) (voxel size 1 mm3, TE = 4 ms, TR = 9 ms) before and after contrast, axial T2w (voxel size 0.36 × 0.36 × 4 mm, TE = 87 ms, TR = 3396 ms), DSC perfusion (voxel size 1.75 × 1.75 × 4 mm, TE = 40 ms, TR = 1547 ms, Flip Angle = 75°, 80 dynamics) and DTI (TR/TE: 5,000/78 ms, voxel size of 2 2 2 mm3, 32 diffusion gradient directions, b value 1,000 s/mm2).

All preoperative MRI scans were uniformly processed: Initially, images were co-registered to the SRI24 atlas space [19] using NiftyReg [20] and then skull-stripped with HD-BET [21]. Tumor segmentation was automated using the BraTS Toolkit [22], which was validated in prior studies [23], segmenting tumors into areas of necrosis, contrast-enhancing tumor, and edema/ non-contrast-enhancing tumor. In cases where T2w or FLAIR images were missing, a GAN-based approach was employed to generate the missing sequences to allow automatic segmentation [24]. Resulting segmentations were quality-checked by M.G. ADC and FA maps were derived from DTI using the dipy library [25]. For leakage-corrected CBV calculations, the method by Arzanforoosh et al. was used, which required a mask of normal-appearing white matter for normalization [26]. ANTs Atropos was utilized for this purpose, excluding tumor regions to maintain accuracy [27]. Subsequently, tumor volumes were automatically extracted from segmentation masks. The resulting segmentations were visually quality checked by M.G. and corrected where necessary. Afterwards, summary statistics for ADC and CBV values were compiled.

Free water correction of preoperative DTI data was conducted using an artificial neural network (ANN) model, as described in a previous study [28]. This ANN was trained with synthetically generated data that included known ground truth, enabling it to learn a nonparametric forward model. This model maps free-water partial volume contamination to volume fractions, effectively decomposing the measured diffusion signal into a “true” diffusion signal and free-water contamination. The model is publicly available at https://github.com/mmromero/dry. Notably, since ADC relies on isotropic diffusion, we only calculated FA maps from free-water corrected DTI data, as well as voxel-wise tissue volume maps, which are essentially given as \(\:tvm\:=\:1\:-\:free\:water\:fraction\).

To automatically segment neurogenic zones in the preoperative MRI, we adopted the methodology developed by Bruil et al. [29] and explained in [10]. In brief, we employed ANTs with parameters established as a robust baseline for the BraTSReg challenge to deformably register the SRI atlas onto each patient’s preoperative MR image. Subsequently, we warped the SVZ atlas from Bruil et al. and the Julich brain atlas for DG segmentation into the patient’s anatomy [29, 30]. All registrations underwent rigorous visual quality checks (by MG) to ensure accuracy. Following the automated tumor and atlas segmentation, we calculated the tumor’s center of mass (of the tumor core, including contrast-enhancing tumor and necrosis) using SciPy like previously described by Jung et al. [10]. The minimum Euclidean distance from this center to the surface of the respective neurogenic zones (SVZ and SGZ) was then automatically determined using SciPy distance functions. This approach ensures accurate and consistent distance calculations, even when dealing with complex structures like the SVZ. This methodology provides a robust framework for quantitatively assessing the spatial relationships between glioblastomas and neurogenic zones. For a detailed visual representation of this process, please refer to Figs. 1 and 2 in our previous publication [10]. The percentage of tumor infiltration is represented by the overlap of segmentation of tumor core and SVZ.

Following formalin fixation and paraffin embedding, tissue samples underwent standard neuropathological diagnosis. Beyond traditional histological and immunohistochemical techniques, an 850k methylation analysis was performed on the extracted DNA using the Illumina EPIC 850k Methylation Array BeadChip (Illumina Inc., San Diego, CA, USA). The methylation data was analyzed utilizing the Brain Tumor Classifier from the German Cancer Research Center (DKFZ) and the University of Heidelberg [31, 32]. This comprehensive approach culminated in an integrated diagnosis that combined histology, immunohistochemistry, and molecular pathology, adhering to the guidelines of the 2021 5th edition of the WHO classification [18].

All statistical analyses were conducted using Python 3 with Matplotlib, Scipy, Seaborn, and Statsmodels libraries. The study aimed to investigate the relationships between tumor infiltration into the SVZ and various metrics, including TVM, CBV and FA-FWE.

Quantitative metrics were first extracted and normalized to compute percentiles (P5, P25, P50, P75, P95) for TVM, CBV, and FA-FWE. Pearson correlation coefficients were used to assess the relationships between SVZ infiltration and the metrics. Independent T-tests were employed to compare the average values of TVM, CBV, and FA-FWE between groups with different levels of SVZ infiltration. Additionally, Pearson correlation coefficients were computed to evaluate the relationship between the minimum distance from the tumor center of mass to the SVZ and the various metrics. Visualizations with scatter blots were used to illustrate the relationships between SVZ infiltration and the metrics.