AI Model Predicts Glioblastoma Infiltration Beyond Visible Tumor Margins

Mapping the Invisible Margins of Glioblastoma Infiltration

Glioblastoma remains the most lethal primary brain malignancy, characterized by a median survival of 12 to 15 months and nearly universal recurrence [1, 2]. While maximal safe resection and chemoradiotherapy are standard, treatment failure often stems from microscopic tumor cells that infiltrate the brain parenchyma well beyond the contrast-enhancing margins visible on standard MRI [3]. Current imaging techniques frequently struggle to distinguish between true tumor progression and pseudoprogression (treatment-related inflammatory changes that mimic tumor growth), a diagnostic challenge where artificial intelligence algorithms have demonstrated a pooled sensitivity of 88% (95% CI, 77% to 100%) in high-grade gliomas [4, 5, 6]. Although advanced sequences such as diffusion-weighted imaging (an MRI technique that measures the random motion of water molecules to infer tissue cellularity) offer insights into intratumoral heterogeneity, they have not yet solved the problem of mapping diffuse infiltration for surgical planning [3]. The SupraGlio trial recently evaluated GlioMap, a computational model that identifies high-risk regions of recurrence before they become visible on conventional scans, achieving an accuracy of 0.81 (95% CI, 0.71 to 0.91) and an AUC of 0.84 for histologically confirmed infiltration [7].

Biopsy-Controlled Validation of Recurrence Risk

The SupraGlio trial (NCT05735171) provides a prospective validation of GlioMap, an open-access artificial intelligence model designed to address the limitations of conventional imaging. This tool utilizes multiparametric MRI data to predict voxelwise infiltration (the identification of tumor cells within individual three-dimensional pixels of an imaging volume) and the subsequent risk of recurrence. To test these predictions, researchers enrolled patients with newly diagnosed glioblastoma to undergo neuronavigated biopsies, a technique using computer-assisted guidance to sample specific brain coordinates with millimeter precision. These biopsies targeted two distinct areas beyond the visible contrast-enhancing tumor margin: high-risk of recurrence (HRoR) regions and low-risk of recurrence (LRoR) regions as identified by the algorithm. This approach allowed the researchers to correlate imaging signatures directly with tissue-level pathology in areas that would typically appear uninvolved on a standard T1-weighted contrast scan.

To establish a definitive baseline for the model's predictions, the researchers used histopathological infiltration as the ground truth, confirming the presence or absence of tumor cells through microscopic tissue examination. The histological validation phase involved the analysis of fifty-eight biopsies from 27 patients. The performance of the model was evaluated using accuracy and the area under the receiver operating characteristic curve (AUC), a statistical measure of a diagnostic test's ability to discriminate between diseased and non-diseased states. GlioMap achieved an accuracy of 0.81 (95% CI, 0.71 to 0.91) and an AUC of 0.84 (95% CI, 0.73 to 0.93) for identifying histologically confirmed tumor infiltration in these peritumoral regions. For the clinician, these metrics suggest a high degree of reliability in identifying occult disease that currently evades standard surgical and radiotherapeutic margins.

Molecular Evidence of a Neural-to-Mesenchymal Gradient

To confirm that the AI model's spatial predictions reflected the underlying biology of the disease, the researchers conducted a multi-layered validation involving histopathological assessment, transcriptomic profiling, and survival analysis. Transcriptomic profiling (the measurement of all messenger RNA molecules within a tissue sample to characterize its gene expression or molecular phenotype) was performed on 48 samples obtained from 16 patients. This analysis allowed the team to correlate the AI-predicted risk levels with the actual genetic activity of the cells in those specific coordinates beyond the visible tumor margin, providing a molecular map of the infiltrative front. This is particularly relevant as glioblastoma cells often adopt different metabolic and signaling states as they migrate into healthy brain tissue.

The molecular data revealed a clear biological continuum that aligned with the model's risk stratification. The researchers identified a progressive upregulation of genes associated with tumor invasion and angiogenesis, specifically CD44, CHI3L1, STAT3, and VEGFA, as they moved from low-risk of recurrence (LRoR) regions to high-risk of recurrence (HRoR) regions and finally into the tumor core. In tandem with the increase in aggressive markers, the analysis showed a significant downregulation of neuronal markers, including MBP (myelin basic protein) and GABRA1 (a gamma-aminobutyric acid receptor subunit), across the same spatial transition from LRoR to the tumor center.

These distinct gene expression patterns confirmed the presence of a neural-to-mesenchymal gradient, which is a transition from the characteristics of healthy brain tissue to the aggressive, mesenchymal traits associated with glioblastoma infiltration and poor prognosis. For the practicing clinician, this molecular evidence suggests that the AI tool is accurately mapping the biological front of the malignancy. By identifying HRoR zones where neuronal markers are lost and pro-angiogenic factors are increased, the model provides a biologically grounded target for potentially expanding surgical margins or intensifying radiation therapy in areas that appear normal on standard imaging.

Prognostic Value of Residual High-Risk Volume

The clinical utility of the GlioMap model rests on its ability to address the primary cause of treatment failure: glioblastoma recurrence is driven by diffuse microscopic infiltration beyond the contrast-enhancing tumor margin. To determine if the AI-identified regions correlated with patient outcomes, the researchers conducted survival analyses that assessed the prognostic value of postoperative high-risk of recurrence (HRoR) volume. This metric represents the amount of tissue the model identifies as highly infiltrated that remains in the brain following surgical resection. By quantifying this residual burden, clinicians can better understand the likelihood of early recurrence even when standard postoperative imaging suggests a gross total resection of the visible tumor mass, which often provides a false sense of security regarding local control.

The study established a specific threshold for this residual infiltration that significantly impacts patient longevity and disease control. The researchers found that a postoperative HRoR volume greater than 1.6 cubic centimeters predicted shorter overall survival (P = .04). Furthermore, this same volume threshold was an even stronger predictor of disease recurrence, as a postoperative HRoR volume greater than 1.6 cubic centimeters predicted shorter progression-free survival (P = .008). These findings indicate that the volume of microscopic disease left behind, as mapped by the AI, serves as a critical determinant of the clinical course, independent of the visible tumor margins seen on conventional MRI.

By accurately identifying histologically and transcriptionally infiltrated regions, GlioMap provides a biologically grounded imaging biomarker (a measurable indicator of a biological state or condition) that could fundamentally alter postoperative management. For the practicing neurosurgeon and radiation oncologist, these data suggest that the model could be used to guide extended resection or to refine and personalize radiotherapy planning. Instead of applying uniform margins, clinicians might use these AI-generated maps to intensify treatment in specific high-risk zones, potentially improving local tumor control and extending survival for patients with this aggressive malignancy.

References

1. Poursaeed R, Mohammadzadeh M, Safaei AA. Survival prediction of glioblastoma patients using machine learning and deep learning: a systematic review.. BMC cancer. 2024. doi:10.1186/s12885-024-13320-4

2. Wen PY, Weller M, Lee EQ, et al. Glioblastoma in adults: a Society for Neuro-Oncology (SNO) and European Society of Neuro-Oncology (EANO) consensus review on current management and future directions. Neuro-Oncology. 2020. doi:10.1093/neuonc/noaa106

3. Brancato V, Nuzzo S, Tramontano L, Condorelli G, Salvatore M, Cavaliere C. Predicting Survival in Glioblastoma Patients Using Diffusion MR Imaging Metrics—A Systematic Review. Cancers. 2020. doi:10.3390/cancers12102858

4. Palavani LB, Nogueira BV, Mitre LP, et al. Artificial intelligence algorithms for differentiating pseudoprogression from true progression in high-grade gliomas: A systematic review and meta-analysis.. Neurosurgical review. 2025. doi:10.1007/s10143-025-03718-4

5. Booth TC, Grzeda M, Chelliah A, et al. Imaging Biomarkers of Glioblastoma Treatment Response: A Systematic Review and Meta-Analysis of Recent Machine Learning Studies.. Frontiers in oncology. 2022. doi:10.3389/fonc.2022.799662

6. Sidibe I, Tensaouti F, Roques M, Cohen-Jonathan-Moyal E, Laprie A. Pseudoprogression in Glioblastoma: Role of Metabolic and Functional MRI-Systematic Review.. Biomedicines. 2022. doi:10.3390/biomedicines10020285

7. Cepeda S, Hernando-Pérez E, Pérez-Riesgo E, et al. Prospective biopsy-controlled validation of an AI model for predicting glioblastoma infiltration: Results from the SupraGlio trial.. Neuro-oncology. 2026. doi:10.1093/neuonc/noag088