Research Associate University of California San Francisco
Disclosure(s):
Katie Scotford, BA: No financial relationships to disclose
Introduction: It is well established that cytoreduction improves overall (OS) and progression free survival (PFS) for patients with diffuse glioma. We recently demonstrated an OS benefit for glioma resection outside of the imaging defined tumor margins in subsets of patients with low and high-grade disease. However, current surgical strategies lack the ability to identify infiltrative tumor cells at the margin. In this study we applied histological and optical imaging of glioma specimen using a patch based Stimulated Raman histology (SRH) platform.
Methods: We acquired 714 distinct glioma specimens for SRH including both tumor core and margin samples for model development including AI generated indications of glioma burden across two institutions for both training and validation. Each tumor sample was scored histologically using tumor specific tissue characterization for generalized linear mixed models with agreement established between neuropathology and AI generated values.
Results: Histologically we confirmed that dense glioma within each sample was correctly classified with high accuracy. At the sample level we reported an AUROC of 91.2% for dense tumor using H&E and an AUROC of 95.8% for dense tumor using SRH. We next established a convolutional neural network (CNN) in order to increase classifier accuracy and report an AUROC of 96.8% for dense tumor detection at the infiltrative margin using SRH.
Conclusion : We demonstrated excellent accuracy for glioma detection including samples obtained beyond the imaging defined tumor margins. These results support the opportunity to introduce a non-inferiority clinical trial supporting the prospective use of a CNN built for SRH compared with the diagnostic accuracy of a trained neuropathologist.