Abstract
AbstractGlioma are often impossible to visualize discrimination within different grades and staging, especially for glioma molecular subgrouping which is highly related with surgery strategy and prognosis. Based on glioma guideline published on 2021, molecular subgroups such as IDH, 1p/19q etc. need to be detected to classify the subgroups (astrocytoma, oligodendroglioma, GBM) from high-grade glioma and guide the personalized treatment. However, timely intraoperative technology is limited to identify molecular subgroups of glioma tissues. To address this problem, we develop a deep learning-guided fiberoptic Raman diagnostic platform to assess its ability of real-time high-grade glioma molecular subgrouping. The robust Raman diagnostic platform is established using convolutional neural networks (ResNet) together with fingerprint spectra acquired within 3 seconds. We have acquired a total of 2358 Raman spectra from 743 tissue sites (astrocytoma: 151; oligodendroglioma:150; GBM: 442) of 44 high-grade glioma patients (anaplastic astrocytoma: 7; anaplastic oligodendroglioma:8; GBM: 29). The optimized ResNet model provides an overall mean diagnostic accuracy of 84.1% (sensitivity of 87.1% and specificity of 81.5%) for identifying 7 molecular subgroups (e.g., IDH, 1p/19q, MGMT, TERT, EGFR, Chromosome 7/10, CDKN2A/B) of high-grade glioma, which is superior to the best diagnosis performance using PCA-SVM and UMAP. We further investigate the saliency map of the best ResNet models using the correctly predicted Raman spectra. The specific Raman features that are related to the tumor-associated biomolecules (e.g., collagens, and lipids) validate the robustness of ResNet diagnostic model. This potential intraoperative technology may therefore be able to diagnosis molecular subgroups of high-grade glioma in real time, making it an ideal guide for surgical resection and instant post-operative decision-making.
Publisher
Cold Spring Harbor Laboratory