Objective:

To assess whether AI analysis of intraoperative tissue samples, combined with clinical, radiographic, and molecular characteristics, can predict focal glioblastoma recurrence near sampled resection margins.

Approach:

• Study design: Multicenter retrospective study of adults with IDH-wild-type glioblastoma who underwent resection and stimulated Raman histology (SRH), followed by a separate prospective internal validation cohort.
• Model development: Researchers analyzed 273 margin samples from 60 patients using FastGlioma to generate tumor-infiltration scores, then evaluated six machine-learning classifiers with 10-fold internal cross-validation.
• Validation: The selected model was tested in 20 additional patients from the same health system. Recurrence locations were assessed using follow-up imaging and compared with sampled margin sites.

Key Findings:

• AI-derived infiltration scores were higher at sites that later recurred than at nonrecurrent sites (median 0.74 vs 0.60).
• AI infiltration alone had a mean AUC of 0.62; histopathologic infiltration alone had a mean AUC of 0.59.
• A random forest model had a mean AUC of 0.87 in internal cross-validation and an AUC of 0.80 in the separate internal validation cohort.
• A reduced model using AI-derived infiltration, age, and resection cavity volume had a mean cross-validation AUC of 0.82.
• In spatial cross-validation analyses, the model’s mean AUC was 0.80 for recurrence within 5 mm and 0.76 within 10 mm of a sampled site.
• SRH imaging of unprocessed tissue took less than 60 seconds per sample.

Interpretation:

The authors reported that combining optical assessment of tumor infiltration with clinical variables may help estimate locations at risk for focal recurrence before recurrence is visible on standard imaging.

Limitations:

• The number and distribution of margin samples were not standardized, and infiltration between sampled locations could not be characterized.
• Sampling was constrained by functional mapping and proximity to eloquent brain regions.
• All patients underwent maximal resection; performance in patients with unresectable glioblastoma is unknown.
• The validation cohort came from the same health system as the development cohort.
• The 5-mm and 10-mm spatial estimates came from cross-validation, not the separate validation cohort.

Conclusion:

The study found that AI-assisted SRH combined with clinical characteristics predicted focal recurrence in an internal validation cohort; further validation beyond the same health system was not reported.

Sources:

• Science Advances (article source named in provided context)