IMG-29. Supratotal Insights: MRI-Based Deep Radiomic Features from Tumor Margins for Overall Survival (OS) Prediction in Glioblastoma Multiforme (GBM) Patients
Year: 2025
Type: conference-abstract
Abstract: GBM is a highly heterogeneous brain tumor with limited survival outcomes. While conventional radiomic analyses predominantly target the tumor, recent clinical methods like supratotal resection highlight the prognostic relevance of the peritumoral region. This study investigates whether deep radiomic features derived from peritumoral MRI zones can improve OS prediction. Multimodal MRI data (T1, T1Gd, T2, FLAIR) from 520 GBM patients were utilized. Tumor and peritumoral regions were pre-processed and segmented by performing progressive morphological dilations beyond the whole tumor (Whole Tumor = Tumor core + Edema + Necrosis) boundary. Radiomic features were extracted using deep learning models which includes ViT-B16 and ResNet50, along with clinical data. A feature selection pipeline was applied to identify the most relevant radiomic signatures which results in 180 features. Support Vector Machine classifier were trained and validated to classify short survivors (≤12 months) and long survivors (>12 months) to evaluate the predictive contribution of tumor and peritumoral regions. The baseline survival prediction using tumor-only features (0 mm margin) achieved 10-fold cross validated AUC of 0.74 ± 0.07 (95% CI: 0.69–0.80). As peritumoral regions were progressively included, a gradual improvement in predictive performance was observed. The highest performance was recorded at the 2.1 mm margin with whole tumor, yielding a cross validated AUC of 0.76 ± 0.06 (95% CI: 0.71–0.80). Notably, peritumoral regions alone also demonstrated promising predictive value, with performance increasing at larger margins. Across all margins, models combining tumor and peritumoral features consistently outperformed those using peritumoral regions alone. These results emphasize the critical prognostic value of the immediate peritumoral zone in improving non-invasive OS prediction for GBM. Including peritumoral features specifically 2.1 mm margin, significantly enhances GBM survival prediction, shows the prognostic importance of tumor-adjacent regions in non-invasive modeling and supports the relevance of emerging surgical strategies like supratotal resection for advancing personalized patient care.
Source: Neuro-Oncology
Institution: International Institute of Information Technology
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4416140100
Cites:
Cited by:
Related to: