Related Experiment Video
Updated: Sep 4, 2026
Digital Spatial Profiling for Characterization of the Microenvironment in Adult-Type Diffusely Infiltrating Glioma
Published on: September 13, 2022
Radiomics-Inferred Immune Programs Map Therapeutic Contexts in Glioblastoma
Daisuke Kawahara1, Misato Kishi1, Yuzuha Kadooka1
1Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences, Hiroshima University, Hiroshima 734-8551, Japan.
Rationale And Objectives:
To develop and evaluate a noninvasive radiogenomic framework linking multiregional magnetic resonance imaging radiomics to glioblastoma immune microenvironment states and mechanism-based therapeutic hypotheses.
Materials And Methods:
The internal University of California San Francisco-Preoperative Diffuse Glioma MRI dataset (UCSF-PDGM) cohort included 223 patients, divided into training (n = 156) and validation (n = 67) sets; 98 The Cancer Genome Atlas glioblastoma dataset (TCGA-GBM) patients formed an independent external cohort. A block/atom-based ensemble radiomics model was developed for overall survival prediction. Matched bulk RNA sequencing from 88 external-cohort patients underwent non-negative least-squares deconvolution using a GSE84465-derived single-cell reference, pathway analysis, and single-cell-supported interpretation. Exploratory Hot, Warm, and Cold states were defined by K-means clustering of direct T-cell-marker expression and interferon-gamma-response single-sample gene set enrichment analysis scores. Therapeutic candidates were organized by immune state and mechanism without assigning efficacy scores.
Results:
The radiomics model achieved C-indices of 0.787 (95% confidence interval [CI], 0.741-0.826), 0.701 (95% CI, 0.612-0.777), and 0.702 (95% CI, 0.638-0.760) in the training, validation, and external cohorts, respectively; corresponding mean time-dependent areas under the curve were 0.857, 0.773, and 0.805. The ensemble prognostic score remained independently associated with overall survival in all cohorts. In external testing, the radiomics-plus-clinical model showed a higher C-index than the clinical-only model (0.730 vs 0.639). Transcriptomic analyses indicated heterogeneous tumor-microenvironment composition and compartment-restricted immune programs. The final classification comprised 12 Hot (13.6%), 23 Warm (26.1%), and 53 Cold (60.2%) tumors and showed bootstrap stability (adjusted Rand index, 0.86). Immune-state-specific therapeutic assignments were interpreted as hypothesis-generating.
Conclusions:
This framework links magnetic resonance imaging-derived prognostic phenotypes to exploratory immune states and mechanism-based therapeutic hypotheses in glioblastoma. Independent radiogenomic and prospective validation is required before clinical application.
