P04.08.A NEAR REAL-TIME INTRAOPERATIVE CLASSIFICATION OF SPINAL TUMORS USING LABEL-FREE STIMULATED RAMAN HISTOLOGY AND ARTIFICIAL INTELLIGENCE

Abstract BACKGROUND Accurate intraoperative tissue diagnosis is essential for surgical guidance, preserving neurological function, and selecting appropriate adjuvant therapies in spinal tumor surgery. Current state-of-the-art AI models trained on intracranial CNS tumors achieve low diagnostic performance when applied to spinal tumors, highlighting the need for specialized diagnostic tools. This diagnostic gap leads to reliance on time-consuming frozen sections, creating a critical need for dedicated spinal tumor classification systems. MATERIAL AND METHODS We developed an AI-powered computer vision tool using a portable stimulated Raman histology (SRH) imager that consumes acrylic slides with squeezed fresh unprocessed tissue specimens. To test the model’s performance, we conducted an international multicenter study enrolling patients across three institutions (New York University, University of Michigan, and Medical University of Vienna). Using a SRH-specific foundation model (University of Cologne), pretrained with near whole CNS tumor spectrum and combined with a novel patch-based transformer classifier, we trained the system for common spinal tumor classification and visualization into (1) Meningioma CNS WHO Grade 1-2, (2) Schwannoma CNS WHO Grade 1, (3) Ependymoma spinal, myxopapillary CNS WHO Grade 2, and (4) Metastases of various primary origin. We compared performance against the existing state-of-the-art SRH CNS classification model on our testing cohort. Final integrated neuropathological diagnosis served as ground-truth. RESULTS We analyzed 140 intraoperative SRH slides comprising ependymomas (n=58, 41.4%), metastases (n=45, 32.1%), schwannomas (n=23, 16.4%), and meningiomas (n=14, 10.0%). Our model achieved 90.87% mean class accuracy (95% CI: 87.30-94.33) with 93.58% macro-average AUROC (95% CI: 91.30-95.72) at slide-level, significantly outperforming the current state-of-the-art model (56.89% MCA). Individual class accuracies were: schwannoma 23/23 (100%), meningioma 14/14 (100%), metastasis 41/45 (91.23%), and ependymoma 42/58 (72.24%). The system demonstrated high specificity (95.17%, 95% CI: 93.20-97.01) and sensitivity (90.87%, 95% CI: 87.30-94.33). We observed 16 false negatives and 2 false positives for ependymomas, while metastases showed 4 false negatives and 9 false positives. The system delivered diagnostic results with interpretable visual feedback within 3 minutes of slide processing. CONCLUSION Our model is the first dedicated intraoperative AI vision system for classifying spinal tumors specifically. The performance shift between models demonstrates that domain-specific AI development with transformer-integrated architecture achieves a 30% performance boost. Moreover, this tool enables near real-time surgical decision-making and guidance, potentially reducing reliance on frozen sections.