Clinical Report: AI may predict focal GBM recurrence
Overview
In a study of resected IDH-wildtype glioblastoma, an AI model combining stimulated Raman histology (SRH)-derived tumor infiltration with clinical, radiographic, and molecular variables predicted focal recurrence at sampled margin sites. The random forest model achieved an AUC of 0.87 in internal cross-validation and 0.80 in a separate 20-patient internal validation cohort; the validation cohort came from the same health system as the development cohort.
Background
Glioblastoma is an aggressive brain tumor with a high rate of recurrence, most often near the residual tumor or resection cavity, although distant recurrence can occur. The study focused on whether microscopic infiltration in tissue collected during resection could help identify margin locations at risk for later focal recurrence. SRH enabled imaging of fresh, unprocessed tissue, and FastGlioma generated an infiltration score for each sample. The investigators compared sampled locations with subsequent contrast-enhancing recurrence on imaging assessed using RANO 2.0 criteria.
Data Highlights
| Measure | Result | Analysis context |
|---|---|---|
| Development cohort | 60 patients; 273 margin samples; 43 patients experienced recurrence | University of California, San Francisco |
| Separate internal validation cohort | 20 patients; 98 margin samples | Same health system as development cohort |
| Median AI infiltration score | 0.74 at recurrent sites vs 0.60 at nonrecurrent sites | Sampled margin sites |
| AI infiltration alone | Mean AUC 0.62 | Recurrence prediction |
| Histopathologic infiltration alone | Mean AUC 0.59 | Recurrence prediction |
| Random forest model | Mean AUC 0.87 | 10-fold internal cross-validation |
| Random forest model | AUC 0.80 | Separate internal validation cohort |
| Reduced random forest model | Mean AUC 0.82 | 10-fold internal cross-validation; AI infiltration, age, and resection cavity volume |
| Spatial recurrence prediction | Mean AUC 0.80 within 5 mm; 0.76 within 10 mm | 10-fold internal cross-validation |
| SRH imaging time | Less than 60 seconds per sample | Fresh, unprocessed tissue |
Key Findings
- AI-derived infiltration scores were higher at margin sites where tumors subsequently recurred than at nonrecurrent sites (median 0.74 vs 0.60); histopathologic assessment also showed greater infiltration at recurrent sites.
- Infiltration alone had limited predictive performance: mean AUC was 0.62 for AI-derived infiltration and 0.59 for histopathologic infiltration.
- Among six classifiers, the random forest model performed best, with a mean AUC of 0.87 in 10-fold internal cross-validation and an AUC of 0.80 in the separate 20-patient internal validation cohort.
- AI-derived infiltration was the leading predictor in five of six classifiers and remained the leading predictor in a reduced random forest model containing infiltration, age, and resection cavity volume; that reduced model had a mean cross-validation AUC of 0.82.
- In spatial analyses based on internal cross-validation, the model had mean AUCs of 0.80 for recurrence within 5 mm and 0.76 within 10 mm of a sampled margin site; these estimates were not from the separate validation cohort.
- The authors noted limitations including nonstandardized margin sampling, inability to characterize infiltration between sampled locations, sampling constraints near eloquent brain, maximal resection in all cohorts, and validation within the same health system as model development.
Clinical Implications
The authors state that combining optical histology of GBM infiltration with clinical variables may enable estimation of focal recurrence locations before recurrence is visible on standard imaging. The reported performance is based on internal cross-validation and a separate validation cohort from the same health system; the study does not establish performance in patients with unresectable glioblastoma.
Conclusion
SRH-derived infiltration combined with clinical and tumor characteristics showed predictive performance for focal recurrence at sampled resection margins in this study. The authors identify sampling limitations and the same-system validation cohort as constraints on interpretation.
Related Resources & Content
- Herr S, et al., Science Advances, year not provided — AI-assisted optical microscopy of tissue collected during resection of glioblastoma may help predict focal recurrence
- Frontiers in Neurology, 2026 — Machine and deep learning based on magnetic resonance imaging to segment glioblastoma and predict the spread of recurrence: a multicenter retrospective protocol
- Louis DN, et al., Neuro-Oncology, 2021 — The 2021 WHO Classification of Tumors of the Central Nervous System: a summary
- ASCO Publications — Tumor recurrence or treatment effect? Large multi-institutional evaluation of an AI risk assessment model in glioma and brain metastases.
- ASCO Publications — Tumor recurrence or treatment effect? Large multi-institutional evaluation of an AI risk assessment model in glioma and brain metastases.
- ASCO Publications — AI–assisted survival prediction in glioma: Meta-analysis of prospective and controlled studies.
- Machine and deep learning based on magnetic resonance imaging to segment glioblastoma and predict the spread of recurrence: a multicenter retrospective protocol
- The 2021 WHO Classification of Tumors of the Central Nervous System: a summary
- 2021 WHO Classification of Tumors of the Central Nervous System: a summary | Neuro-Oncology | Oxford Academic
- MGMT Gene Silencing and Benefit from Temozolomide in Glioblastoma | New England Journal of Medicine
- Spatial patterns of progression in reported glioblastoma cohorts after upfront chemoradiation and salvage therapy: a systematic review and meta-analysis.
This content is an AI-generated, fully rewritten summary based on a published scholarly article. It does not reproduce the original text and is not a substitute for the original publication. Readers are encouraged to consult the source for full context, data, and methodology.
