Predicting Glioma Survival From Intratumoral Vascular Topology: A Connectional Vessel Template Approach.
Authors
Affiliations (8)
Affiliations (8)
- Institute of Biomedical Engineering, Bogazici University, Istanbul, Turkey.
- Basaksehir Cam and Sakura City Hospital, Istanbul, Turkey.
- Department of Medical Pathology, Acibadem University, Istanbul, Turkey.
- Center for Neuroradiological Applications and Research, Acibadem University, Istanbul, Turkey.
- Electrical and Electronics Engineering Department, Systems Science and Mathematics Laboratory, Medical Magnetic Devices Laboratory, Bogazici University, Istanbul, Turkey.
- Department of Neurosurgery, Acibadem University, Istanbul, Turkey.
- Department of Radiology, Acibadem University, Istanbul, Turkey.
- Bogazici University Center for Targeted Therapy Technologies, Istanbul, Türkiye.
Abstract
Gliomas exhibit aberrant angiogenesis that drives tumor progression and therapeutic resistance. Susceptibility-weighted imaging (SWI) enables noncontrast visualization of intratumoral vasculature, but current approaches rely on summary metrics that overlook vascular topology. We developed a graph-based framework to model vascular networks from SWI and predict glioma survival. Between 2005 and 2020, 228 adults with histologically confirmed diffuse glioma (World Health Organization [WHO] Grades 2-4) who had undergone preoperative MRI were retrospectively screened. MRI was performed on 1.5- and 3-T scanners (Siemens Healthineers, Erlangen, Germany) including T2-weighted imaging (TR/TE 4000-6000/80-120 ms, FOV 220-256 mm, slice thickness 2-3 mm) and SWI (TR/TE 25-35/15-25 ms, FOV 200-240 mm, slice thickness 1.5-2 mm). Tumors were manually delineated on T2-weighted images and coregistered to SWI using advanced normalization tools (ANTs). Intratumoral vessels were extracted from SWI using the Frangi vesselness filtering and represented as graphs with nodes at branch points and endpoints and edges along vessel segments. Node-level, edge-level, and graph-level vascular features were processed by a GraphSAGE encoder to generate patient-specific embeddings, which were compared to a learnable population template within the connectional vessel template (CVT) framework to quantify individual vascular atypicality. A Cox proportional hazards head integrated graph embeddings, template deviation, and demographic features to output risk scores. Model performance was evaluated using 10-fold stratified cross-validation for progression-free survival (PFS) and overall survival (OS), with multivariable Cox regression and subgroup analyses across isocitrate dehydrogenase (IDH) mutation status and histological subtypes. A total of 193 patients were included (mean age 43.4 ± 14.0 years; 57.5% male). OS analysis included 182 patients (61 deaths), and PFS analysis included 170 patients (102 progression events). The CVT model achieved C-indices of 0.812 for PFS and 0.839 for OS with significant risk stratification (PFS log-rank p = 5.82 × 10<sup>-7</sup>; OS log-rank p = 1.17 × 10<sup>-6</sup>). The CVT risk score independently predicted PFS (HR 1.77, p < 0.001) and OS (HR 1.44, p < 0.001) beyond age and sex. Subgroup analyses demonstrated consistent PFS discrimination in IDH-wildtype tumors and glioblastoma (C-index 0.677) and strong OS prediction in IDH-mutant tumors (C-index 0.850). Graph-based modeling of SWI-derived vascular topology provides robust, independent prognostic information for glioma survival using a single noncontrast MRI sequence.