PHSP-Net: Personalized Habitat-Aware Deep Learning for Multi-Center Glioblastoma Survival Prediction Using Multiparametric MRI.

Abstract

Background: Glioblastoma (GBM) is a highly aggressive and heterogeneous primary malignancy of the central nervous system, with a median overall survival (OS) of approximately 15 months. Achieving accurate and generalizable OS prediction across multi-center settings is essential for clinical application.

Methods: We propose a Personalized Habitat-aware Survival Prediction Network (PHSP-Net) that integrates multiparametric MRI with an adaptive habitat partitioning strategy. The network combines deep convolutional feature extraction and interpretable visualization modules to perform patient-specific subregional segmentation and survival prediction. A total of 1084 patients with histologically confirmed WHO grade IV GBM from four centers (UPENN-GBM, UCSF-PDGM, LUMIERE and TCGA-GBM) were included. PHSP-Net was compared with conventional radiomics, habitat imaging models and ResNet10, with independent validation on two external cohorts.

Results: PHSP-Net achieved an AUROC of 0.795 (95% CI: 0.731-0.852) in the internal validation set, and 0.707 and 0.726 in the LUMIERE and TCGA-GBM external test sets, respectively-outperforming both comparison models. Kaplan-Meier analysis revealed significant OS differences between predicted high- and low-risk groups (log-rank p < 0.05). Visualization analysis indicated that necrotic-region habitats were key prognostic indicators.

Conclusions: PHSP-Net demonstrates high predictive accuracy, robust cross-center generalization and improved interpretability in multi-center GBM cohorts. By enabling personalized habitat visualization, it offers a promising non-invasive tool for prognostic assessment and individualized clinical decision making in GBM.