MRI Radiomics Model Predicts IL13Rα2 Expression and Survival in Glioblastoma

Glioblastoma remains one of the most formidable opponents in modern oncology. This aggressive brain tumor, classified by the World Health Organization as a grade IV glioma, carries a grim prognosis despite maximal surgical resection, radiotherapy, and chemotherapy. Median survival rarely exceeds two years, and the molecular heterogeneity of the disease means that patients with seemingly identical tumors can follow dramatically different clinical courses. A study published in BMC Cancer by Yusong Zhang, Jiajun He, and colleagues from The Second Affiliated Hospital of Wenzhou Medical University now reports a way to extract critical molecular information from something clinicians already have in abundance before surgery: the contrast-enhanced T1-weighted magnetic resonance imaging scan.

The target of the investigation is interleukin-13 receptor alpha 2, or IL13Rα2, a cell surface protein that has long intrigued glioblastoma researchers. Unlike the conventional interleukin-13 receptor, IL13Rα2 binds interleukin-13 with high affinity but is generally considered a decoy receptor that also participates in signaling pathways driving tumor growth, invasion, and immune evasion. Prior work has shown that IL13Rα2 is overexpressed in glioblastoma relative to normal brain tissue, and its abundance has been linked to aggressive biological behavior. The new study confirms these associations using transcriptomic data from The Cancer Genome Atlas and the Chinese Glioma Genome Atlas, demonstrating that elevated IL13Rα2 expression is an independent predictor of worse overall survival and is associated with distinctive features of the tumor microenvironment.

What makes IL13Rα2 particularly attractive as a clinical marker is that it is not merely a prognostic indicator. Because the receptor sits on the tumor cell surface and is largely absent from normal brain, it has been explored as a target for antibody-drug conjugates, CAR T cell therapies, and peptide-based immunotherapies. Knowing before an operation whether a patient’s tumor expresses high levels of IL13Rα2 could, in principle, inform therapeutic planning and enrollment into targeted clinical trials. The obstacle has always been measurement: definitive assessment of IL13Rα2 expression currently requires tissue obtained at surgery or biopsy, followed by transcriptomic sequencing or immunohistochemistry. Preoperative prediction has remained out of reach.

The Wenzhou team approached this challenge through radiomics, a computational discipline that converts medical images into high-throughput quantitative data. The underlying premise is that the pixel-level texture, intensity distribution, and spatial patterns visible on a scan reflect, at least in part, the microscopic architecture of the tumor — its cellularity, necrosis, vascularity, and stromal composition. Radiomics analysis begins with segmentation of a region of interest, in this case the contrast-enhancing tumor volume on preoperative CE-T1WI scans. From each segmented region, the researchers extracted 1,197 quantitative features spanning first-order intensity statistics, shape descriptors, and higher-order texture measures derived from matrices such as the gray level dependence matrix, some computed after Laplacian of Gaussian filtering to emphasize structures at particular spatial scales.

Feature stability is a notorious weakness of radiomics studies, so the investigators took careful steps to ensure reproducibility. Two observers independently delineated tumor volumes and extracted features, and only those features with intraclass correlation coefficients exceeding 0.75 — a statistical measure of agreement between raters — were retained. This filtering reduced the feature set to 1,145 stable candidates. The researchers then applied least absolute shrinkage and selection operator regression, a technique that performs feature selection and regularization simultaneously by penalizing the absolute size of regression coefficients, driving irrelevant features toward zero. The resulting composite measure, termed the Rad-score, summarizes the imaging phenotype of each tumor as a single continuous number.

Three machine learning classifiers — logistic regression, random forest, and ExtraTrees — were built on the Rad-score to predict IL13Rα2 expression status. The development cohort consisted of 92 patients from the public datasets who had both paired transcriptomic data and preoperative CE-T1WI scans. The most compelling evidence, however, came from an independent institutional cohort of 36 patients whose CE-T1WI scans could be matched with formalin-fixed paraffin-embedded tissue specimens. In this cross-platform external validation, IL13Rα2 protein abundance was measured by immunohistochemistry, providing a protein-level reference standard rather than relying solely on mRNA expression. The highest-performing model achieved an area under the receiver operating characteristic curve of 0.801 in this institutional cohort, a level of discriminatory accuracy that the authors describe as promising for a noninvasive molecular prediction task.

Beyond predicting receptor expression, the Rad-score carried prognostic weight. Across the combined cohorts, patients with high Rad-scores had significantly shorter overall survival, with a hazard ratio of 2.03 and a 95 percent confidence interval of 1.24 to 3.34, corresponding to a P value of 0.005. In multivariable Cox regression analysis that adjusted for clinical variables, the Rad-score remained significantly associated with overall survival, indicating that the imaging signature contributes prognostic information independent of conventional clinical predictors. Supplementary analyses extended this finding to a nomogram integrating the Rad-score with clinical factors, which achieved area under the curve values of 0.719, 0.797, and 0.728 for predicting 1-, 3-, and 5-year overall survival, respectively, with acceptable calibration across clinically relevant time points.

The study also explored the biological meaning of the imaging signature. The Rad-score was positively correlated with IL13Rα2 expression and with exploratory immune-related features of the tumor microenvironment, including immune cell infiltration patterns estimated by single-sample gene set enrichment analysis. A supplementary scatter plot demonstrated strong concordance between IL13Rα2 mRNA expression Z-scores and protein abundance ratios, with a Spearman correlation coefficient of 0.89 and a Pearson coefficient of 0.84, both highly significant. In the institutional cohort, the continuous Rad-score was moderately and positively correlated with the continuous IL13Rα2 immunoreactive score, with a Spearman coefficient of 0.455 and a P value of 0.005, suggesting that the imaging phenotype captures graded variation in receptor abundance rather than a simple binary classification.

Clinical utility was assessed by decision curve analysis, a framework that evaluates the net benefit of acting on a model’s predictions across a range of probability thresholds, accounting for the relative harms of false positives and false negatives. The models showed net benefit across clinically relevant threshold ranges, supporting their potential as preoperative decision-support tools. The investigators acknowledge that their work is exploratory in several respects. The external validation cohort was modest in size, the cross-platform design — matching mRNA-based training with protein-based validation — introduces an additional layer of biological variability, and the models were developed on CE-T1WI alone rather than multiparametric MRI sequences that might carry complementary information about diffusion and perfusion.

Even with these caveats, the study represents a meaningful step toward preoperative molecular characterization of glioblastoma. If the findings are replicated in larger, prospective, multi-institutional cohorts, a radiomics score computed from a routine preoperative scan could help clinicians stratify patients by molecular risk, guide the timing and extent of surgical intervention, and identify candidates for IL13Rα2-directed therapies before a single incision is made. The convergence of quantitative imaging, machine learning, and cancer immunology on display here illustrates how computational approaches are beginning to bridge the gap between the radiology suite and the molecular pathology laboratory, turning the images clinicians already acquire into a window on the biology that shapes each patient’s fate.

Subject of Research: Preoperative MRI radiomics predicts IL13Rα2 expression and survival in glioblastoma

Article Title: Preoperative MRI radiomics predicts IL13Rα2 expression and survival in glioblastoma

Article References: Zhang, Y., He, J., Qin, D., Hao, J., Wu, Z., Zheng, Y., Tang, F., & Lin, J. (2026). Preoperative MRI radiomics predicts IL13Rα2 expression and survival in glioblastoma. BMC Cancer. https://doi.org/10.1186/s12885-026-16861-y

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16861-y

Keywords: Preoperative, radiomics, predicts, IL13R, expression, survival, glioblastoma, scientific research