Leveraging stimulated Raman histology-based cellularity for random forest prediction of glioblastoma recurrence.

• DOI: 10.1200/jco.2025.43.16_suppl.2010
• Publication Date: 2025-05-28T14:16:16Z
• Authors: Sanjeev Herr; Niels Olshausen; Jasleen Kaur; Youssef Sibih; Vardhaan Ambati; Katie Scotford; Thiebaud Picart; Akhil Kondepudi; Anna-Katharina Me...; Melike Pekmezci; Todd Hollon; Albert H. Kim; Shawn L. Hervey-J...

2010

Background: Glioblastoma is a universally fatal diagnosis with extent of resection being one of the most significant predictors of overall and progression-free survival. Most patients eventually experience recurrence, with sixty percent recurring along the resection cavity. Recent work leveraging Stimulated Raman histology (SRH) and artificial intelligence (AI) has approximated glioma cellularity within the infiltrative margins. It remains unknown if these estimates of glioma burden at the infiltrative margins influence glioblastoma recurrence. This study aims to evaluate a predictive model of focal recurrence in patients with glioblastoma using SRH and AI-generated cellularity scores from tissue samples taken at the resection cavity margins. Methods: A multi-center, retrospective cohort study was conducted on patients diagnosed with glioblastoma who underwent resection followed by spatial annotated tissues acquired from the resection cavity margins. Tissues were analyzed using SRH optical imaging, and histopathology analysis was performed using confocal microscopy. Tissue cellularity was measured histologically and by optical imaging. Results: Over 400 patients and 2,200 specimens were analyzed, of which a nested subset of 60 patients were selected based on selection criteria. Using preoperative and postoperative imaging, margin samples were determined to be in an area of recurrence (n=58) or nonrecurrence (n=220). Cellularity was significantly higher in the recurrent margin sample group when compared to the nonrecurrent group (p = 0.026), which was further confirmed by a pathologist-determined cellularity score (0-3) that demonstrated similar findings (p = 0.026). Results were validated across three medical centers. Six classifiers were then trained for recurrence prediction. Using nineteen of the most predictive variables, random forests (RF) performed best with an AUC of 0.848. RF screening for the minimum practical number of variables demonstrated an AUC of 0.805 using only FastGlioma, age and extent of resection as variables. Conclusions: AI-generated cellularity scores have the potential to predict focal recurrence of glioblastoma, allowing for more tailored approaches to surgical resection and radiotherapy to increase progression-free survival.