Intraoperative classification of CNS lymphoma and glioblastoma by AI-based analysis of Stimulated Raman Histology (SRH).

Introduction: Early diagnosis is important to differentiate central nervous system lymphomas (CNSL) from the main differential diagnosis, glioblastoma (GBM), because of different primary treatment modalities for these entities. Due to neurological deficits, diagnostic stereotactic biopsies often need to be performed urgently. In this setting the availability of an intraoperative neuropathological assessment is limited. Research question: This study uses AI-based analysis of Stimulated Raman Histology (SRH) to establish a classifier distinguishing CNSL from glioblastoma in an intraoperative setting. Material and methods: We collected 126 intraoperative SRH images from 40 patients diagnosed with CNSL. These SRH images were divided into patches, measuring 224 x 224 pixels each. Additionally, we used a comparative dataset of 87 SRH images from 31 patients with GBM as a control group to train and validate a neural network based on the CTransPath architecture. Two distinct diagnostic categories were established: "Lymphoma" and "Glioblastoma". Results: Our model demonstrated an accuracy rate of 92.5% in distinguishing between lymphoma and glioblastoma. Analysis of our test dataset showed a sensitivity of 84.2% and a specificity of 100% in the detection of CNSL, demonstrating performance comparable to standard intraoperative histopathological analysis. Discussion and conclusion: The use of AI-driven analysis of SRH images holds promise for intraoperative tissue examination of stereotactic biopsies with suspected CNSL en par with the current gold standard. This study could improve the management of these cases especially in the emergency setting when conventional intraoperative neuropathological evaluation is unavailable.