Preoperative Prognostic Stratification In GBM Patients Using Inverse Distance-Based Proximity To Stem Cell Niches

Abstract

Glioblastoma multiforme (GBM) is the most lethal brain cancer in adults, with a high local recurrence rate despite surgery and chemoradiotherapy. Emerging evidence has linked tumor aggressiveness with the involvement of stem cell niches (SCN) that vary by patient. However, existing SCN research insufficiently quantifies the geometrical relationship between all SCN zones and the tumor for personalized risk assessment and potentially more effective therapy. This study aims to establish such relationship in brain GBM patients using MRI inverse distance map. Preoperative T1 MRI scans and pre-labeled tumor regions of 102 patients from the publicly available BRATS2017 dataset were used in this study. SCN segmentation, including bilateral subventricular zones and subgranular zones, were manually defined in the Montreal Neurological Institute (MNI) brain template. Based on the binary SCN segmentations, a proximity map was generated using the sum of the power 1 inverse distance weightings to all observations in SCNs, normalized between 0 and 1. All MR scans were transformed into the MNI space in Advanced Normalization Tools (ANTs) using affine alignment followed by symmetric normalization based lesion excluded deformable registration. The resulted transformations were propagated to each tumor segmentation to calculate the mean proximity score (PS), whose prognostic capacity was then evaluated using Cox proportional hazards regression with overall survival (OS) and Log-rank tests between sorted and evenly divided high-risk and low-risk groups. As a comparison, the same analyses were conducted on traditional stem cell niches related features, including tumor edge to the ventricle (EV) and center to the ventricle (CV). PS is correlated with OS with a hazardous ratio (HR) of 4.45 and a p-value of 0.0297, indicating a significant inverse relationship between SCN proximity and survival. In contrast, neither EV nor CV is significantly correlated with OS, with HR = 0.9676, p = 0.7483 and HR = 0.9884, p = 0.5660, respectively. Furthermore, the log-rank test for patients stratified by PS reveals a significant difference in survival risk (log-rank p = 0.0474), while stratification using either EV (log-rank p = 0.1486) or CV ((log-rank p = 0.5829) fails to result in significant differences. We introduced a novel inverse distance-based metric, proximity score, to more effectively characterize GBM tumor to SCN zones geometrical relationships. Proximity score outperformed traditional edge or center distance-based measurements in OS prediction and risk stratification. With the improved quantification, PS may be used for better personalized GBM risk assessment and therapy.