Performance of intraoperative AI-assisted Label-Free Stimulated Raman Histology in Recurrent Brain Tumors

Background Therapy-induced tissue alterations complicate intraoperative diagnosis of recurrent CNS tumors. We evaluated stimulated Raman histology (SRH) combined with convolutional neural networks (CNNs) for rapid tumor detection and classification. Methods In this bicentric observational study, fresh tissue from patients undergoing resection or stereotactic biopsy for suspected recurrent CNS tumors was imaged with fiber-laser-based SRH. CNN1 performed binary tumor/non-tumor prediction and CNN2 histotype classification. CNN1 was evaluated against blinded neuropathologist SRH consensus and CNN2 against the final integrated histopathological diagnosis. Results Of 382 acquired SRH images, 262 images from 143 patients were included. CNN1 achieved an AUC of 0.975 (95% CI, 0.95–1.00) for tumor detection. An exploratory cohort-derived tumor probability threshold of ≥ 59 increased specificity from 70.7% to 92.0% while maintaining 95.2% sensitivity and identifying 71.4% of radiation necrosis samples. Among 212 images in treatment subgroup analysis, CNN1 accuracy did not differ significantly across prior-treatment groups (overall 90.1%; p = 0.080). CNN2 accuracy declined from 61.8% after surgery alone to 39.0% after surgery, radiotherapy, and medical treatment (χ²=6.04, p = 0.049). Multimodal therapy was independently associated with lower odds of correct CNN2 prediction (adjusted OR, 0.10 [95% CI, 0.02–0.39]; p = 0.001 and 0.04 [95% CI, 0.01–0.20]; p < 0.001). 120/382 images (31.5%) were of insufficient diagnostic quality; application of the quality control threshold would have rejected 84/120 (70.0%). Conclusions CNN1 maintained high tumor/non-tumor discrimination across treatment subgroups when image quality was sufficient, whereas CNN2 entity-classification performance declined with increasing treatment intensity. Reliable SRH-based diagnosis in recurrent tumors requires validated image-quality control measures and model adaptation to recurrent disease.