P04.09.A AI-ENHANCED STIMULATED RAMAN HISTOLOGY FOR NEAR REAL-TIME PRIMARY CNS LYMPHOMA IDENTIFICATION AND DIFFERENTIATION FROM OTHER CNS LESIONS

Abstract BACKGROUND Distinguishing primary central nervous system lymphoma (PCNSL) from other brain lesions during surgery remains challenging due to overlapping histological features and time constraints. Accurate differentiation is critically important as the surgical and clinical management strategies differ significantly between PCNSL and other CNS entities. This study evaluates the integration of stimulated Raman histology (SRH) with artificial intelligence for rapid intraoperative diagnostic support. MATERIAL AND METHODS Using SRH technology, we generated virtual H&E-like images from fresh tissue specimen within three minutes without conventional processing. We developed RapidLymphoma, a deep learning algorithm utilizing self-supervised learning, trained on 54,000 SRH microscopic images from surgical specimens collected at four international medical centers. The system was designed to detect PCNSL and differentiate it from other CNS entities, with particular focus on high-grade gliomas and metastases. Final histopathological diagnosis served as the ground-truth standard. RESULTS In a prospective multicenter validation cohort (n=160), RapidLymphoma achieved 97.81% balanced accuracy (±0.91) in detecting and differentiating PCNSL from other CNS entities. For PCNSL detection specifically, the diagnostic tool demonstrated superior performance compared to frozen section analysis (100% vs. 77.77%). Additional validation studies focusing on distinguishing PCNSL from IDH-wildtype diffuse gliomas (n=420) and brain metastases (n=59) demonstrated robust performance with balanced accuracy rates of 95.44% (±0.74) and 95.57% (±2.47), respectively. The system generated interpretable heatmaps highlighting diagnostically relevant morphological features, providing visual feedback for surgeons and pathologists during surgery. CONCLUSION Integrating SRH with RapidLymphoma demonstrates reliable performance for rapid intraoperative PCNSL diagnosis and differentiation from challenging entities such as high-grade gliomas and metastases. This approach provides actionable diagnostic information within three minutes, facilitating timely surgical decision-making and treatment planning.