PATH-72. Ultra-rapid single-cell molecular classification of diffuse gliomas using artificial intelligence and stimulated Raman histology

Abstract INTRODUCTION Previously introduced artificial intelligence (AI) models, such as DeepGlioma, use stimulated Raman histology (SRH) to accurately classify adult diffuse gliomas by molecular subtype according to the 2021 WHO classification guidelines. While enabling near-real-time intraoperative molecular classification, these approaches require complex AI architectures and large images, resulting in increased processing times. We sought to perform a proof-of-concept study to determine whether SRH enables accurate molecular classification of gliomas at single-cell resolution while reducing imaging and computational requirements. METHODS Images were collected intraoperatively using a portable SRH microscope with pre-existing tumor and cell detection algorithms applied as preprocessing. We developed a convolutional neural network focusing on cell nuclei using patient-level labels as ground truth. Our architecture employs a ResNet50 backbone with ROI pooling to extract features from individual cells, followed by a multi-class prediction network that distinguishes between three molecular categories: IDH-wildtype glioblastoma, IDH-mutant astrocytoma, and IDH-mutant, 1p/19q-codeleted oligodendroglioma. The model was trained on 4.91M cells from 207 patients (University of Michigan) and tested on 10.23M cells from 436 patients from 4 international institutions. RESULTS The single-cell model (126.03M parameters) had an estimated imaging time of 2.1s and mean prediction time of 20.06s per 0.14x0.14mm patch, compared to DeepGlioma’s (161.83M parameters) imaging time of 105-195s and mean prediction time of 88.70s per 2.5-4.2x2.8-4.7mm slide, representing an up to 12.8-fold increase in processing speed. Our model achieved 90.28±0.06% accuracy differentiating between the three molecular subtypes of glioma (mAUROC: 0.9423), compared to DeepGlioma’s 91.5% accuracy. Class balanced accuracies for the single-cell model were 88.46%, 85.93%, and 81.24% for glioblastoma, oligodendroglioma, and astrocytoma, respectively. CONCLUSION This study demonstrates that SRH imaging effectively differentiates molecular subtypes of glioma at single-cell resolution. Our approach achieves high accuracy with weakly-supervised labeling and imaging data only while decreasing tissue requirements by 99% and processing times by 260s.