AI infiltration scoring from SRH images outperforms molecular markers in predicting glioblastoma recurrence sites

The study analyzed 400 tissue samples from 80 UCSF patients using SRH, a label-free imaging method that generates pixel-level infiltration values in under a minute. The best model distinguished recurrence-prone from recurrence-free sites more accurately than pathology alone, with the AI score emerging as the top predictor in five of six models. Median time to first recurrence was 5.5 months, and most recurrences occurred at or near the resection cavity.

This approach could allow surgeons to extend resection or plan focal radiation and catheter-based drug delivery to high-risk margins before MRI-visible progression. It addresses a core limitation of glioblastoma surgery: invisible infiltrative cells drive near-universal local relapse despite gross-total resection and standard chemoradiation.

Related work in NEJM and Lancet Oncology has shown that extent of resection and residual tumor volume correlate with survival, yet no real-time, spatially resolved recurrence risk map existed until now. The current model was trained on a modest single-center cohort and tested on a held-out set of 20 patients; external validation and prospective outcome data are still required.

Next steps include multi-center prospective trials to test whether targeted interventions based on these predictions extend progression-free survival, plus integration into intraoperative workflows for immediate surgical decision support.