Artificial intelligence–based tumor detection using stimulated Raman histology in patients undergoing neurosurgical biopsies
Abstract
Objective
Intraoperative confirmation of diagnostic tumor tissue is critical in neurosurgical biopsies, particularly in stereotactic procedures, in which sampling error remains a major limitation. Stimulated Raman histology (SRH) enables rapid, label-free tissue imaging directly in the operating room. This study aimed to evaluate the performance of intraoperative SRH combined with artificial intelligence (AI)–based tumor detection using a CE (Conformité Européenne)–certified algorithm under routine clinical conditions, with a focus on scan-level analysis and the impact of tissue heterogeneity.
Methods
The authors conducted a retrospective single-center study including all consecutive patients who underwent neurosurgical biopsy with intraoperative SRH imaging between December 2020 and October 2025. All acquired SRH scans were analyzed using the Glioma Reveal algorithm, which provides a probabilistic analysis of tumor cell presence. Diagnostic performance at the scan level (defined as a specific region of interest [ROI]) was assessed. After SRH analysis, each tissue sample was investigated by routine histopathological analysis. For statistical analysis, all cases were categorized into three groups based on their final histopathological diagnosis: 1) common tumor entities, 2) rare tumor entities, and 3) nonneoplastic lesions. In discordant cases between AI-based classification and histopathological assessment, a board-certified neuropathologist performed a secondary review of the corresponding SRH images to assess whether the AI prediction was morphologically plausible.
Results
Analysis of 204 SRH scans obtained in 89 patients using the Glioma Reveal algorithm showed an initial overall accuracy of 80.9% for tumor tissue detection. After expert review of the misclassified cases, accuracy increased to 94.1%. In the group of gliomas and lymphomas (70 patients, 168 scans), the initial accuracy was 81.5% (sensitivity 78.8%, specificity 100.0%). Expert review confirmed that 80.6% (25 scans) of false-negative results were morphologically correct within the specific ROI, despite global specimen positivity. After correction for these cases, accuracy increased to 96.4%. In rare tumors (8 patients, 20 scans), diagnostic accuracy rose from 90.0% to 100.0% after expert review. In nonneoplastic lesions (11 patients, 16 scans), initial accuracy remained at 62.5% regardless of expert review due to reactive mimics.
Conclusions
SRH imaging using the Glioma Reveal algorithm provides rapid intraoperative confirmation of representative tumor tissue during biopsy. Although automated performance is robust, intraspecimen heterogeneity and spatial sampling differences significantly impact diagnostic metrics and should be taken into account.
