Artificial intelligence-based analysis and diagnosis of intradural extramedullary spinal tumors by stimulated Raman histology.

Background: Intraoperative Stimulated Raman Histology (SRH) has been reported to be fast and accurate in the assessment of neuro-oncological lesions. However, its application to spinal tumors, especially intradural extramedullary tumors (IDEM), remains underexplored. IDEM primarily include meningiomas and schwannomas, as well as less common entities such as metastases or ependymomas. Given that surgical resection is the primary treatment modality, rapid, artificial intelligence (AI)-driven intraoperative tumor classification based on SRH could enhance surgical decision-making and subsequent management. Methods: We acquired 232 SRH images from patients with IDEM using the NIO Laser Imaging System (Invenio Imaging Inc.). We categorized images into three diagnostic classes: "Meningioma," "Schwannoma," and "Other." Images were divided into 224 × 224 pixel patches and used to train and test AI-based image classifiers employing CTransPath, ResNet, and Vision Transformer architectures. Results: Our best-performing model, utilizing the CTransPath architecture, achieved a classification accuracy of 94.3% on the test dataset. Vision Transformer-based models also performed well, exceeding 90% accuracy, while ResNet models attained slightly lower accuracies (79.6%-88.8%). Qualitative analysis indicates that the top-performing model primarily relies on cellular morphology for classification. Conclusions: Our findings confirm the feasibility and effectiveness of AI-assisted SRH analysis for distinguishing IDEM tumor types. This approach may complement conventional intraoperative neuropathology by providing rapid, reliable, and clinically actionable diagnostic information.