Artificial intelligence (AI)–assisted optical microscopy of tissue collected during resection of glioblastoma (GBM)may help predict where tumors will subsequently recur. A machine-learning model combining microscopic tumor infiltration with clinical, radiographic, and molecular characteristics showed an area under the receiver operating characteristic curve of 0.80 in an internal validation cohort.
The investigators initially assembled a multinational, multicenter retrospective cohort of patients older than 18 years with pathology-confirmed isocitrate dehydrogenase–wild-type glioblastoma who underwent tumor resection and intraoperative stimulated Raman histology (SRH). For model development, the authors analyzed 273 margin samples from 60 patients treated at the University of California, San Francisco; 43 patients subsequently experienced recurrence. The model was then tested in a separate prospective internal cohort of 20 patients with 98 margin samples from the same health system. Fresh, unprocessed tissue collected from resection cavity margins was imaged with SRH and analyzed using FastGlioma, which generated an AI–derived tumor infiltration score ranging from 0 to 1.
Investigators evaluated whether AI-derived infiltration combined with routinely available clinical, radiographic, and molecular variables could predict focal recurrence at sampled margin sites. Follow-up imaging was assessed according to Response Assessment in Neuro-Oncology 2.0 criteria, and the authors compared sample locations with sites of new contrast-enhancing recurrence. They evaluated 6 machine-learning classifiers using 10-fold internal cross-validation and separately modeled recurrence within 5 mm and 10 mm of each sampled site.
AI-derived infiltration was greater at margin sites where tumors subsequently recurred, with a median infiltration score of 0.74 compared with 0.60 at nonrecurrent sites. Histopathologic assessment similarly showed greater infiltration at recurrent sites. However, AI-derived infiltration alone had a mean area under the curve (AUC) of 0.62, and histopathologic infiltration alone had a mean AUC of 0.59, prompting the authors to evaluate multivariable models.
Of the 6 classifiers, the random forest model showed the strongest predictive performance, with a mean AUC of 0.87 during 10-fold internal cross-validation. In the separate 20-patient internal validation cohort, the model achieved an AUC of 0.80. AI-derived infiltration was the most predictive variable in 5 of the 6 classifiers, and higher infiltration contributed to a greater predicted probability of focal recurrence in the random forest model.
A reduced random forest model containing 3 variables—AI-derived infiltration, patient age, and resection cavity volume—maintained a mean AUC of 0.82 during internal cross-validation. AI-derived infiltration remained the leading predictor in this reduced model.
In separate spatial analyses, the model achieved a mean AUC of 0.80 for predicting recurrence within 5 mm of a sampled margin site and 0.76 within 10 mm. These estimates were derived from 10-fold internal cross-validation rather than the separate validation cohort. In an illustrative case, a margin sample with an AI-infiltration score of 0.94 corresponded with the location of recurrence identified on subsequent imaging. SRH imaging of unprocessed tissue could be completed in less than 60 seconds per sample.
The study had several limitations. Investigators did not specify a minimum number or standardized distribution of margin samples and could not characterize tumor infiltration between sampled locations. Sampling was also constrained by functional mapping and proximity to eloquent brain regions, limiting uniform sampling across the resection margin. In addition, all patients in the development and validation cohorts underwent maximal resection, leaving the model's performance in patients with unresectable glioblastoma unknown. The separate validation cohort was drawn from the same health system as the development cohort.
The findings suggest that microscopic assessment of tumor infiltration during initial resection, when combined with clinical and tumor characteristics, may help estimate locations at risk for subsequent focal recurrence before recurrence is visible on standard imaging. “Together, these results suggest that optical histology of GBM infiltration, combined with clinical variables, may enable the estimation of focal recurrence before standard-of-care imaging,” wrote lead study author Sanjeev Herr, of Drexel University College of Medicine, and colleagues.
Disclosures: The authors declared no competing interests.
Source: Science Advances
