Intraoperative artificial intelligence for brain tumor surgery: from tissue diagnosis to real-time surgical guidance

Abstract

The 2021 WHO CNS tumor classification marked a fundamental shift in neuro-oncology: molecular markers rather than histological appearance now define glioma diagnosis and guide treatment. Yet, in clinical practice the molecular information that matters most to a surgeon—IDH status, 1p/19q codeletion, and methylation class—typically arrives days to weeks after the operation it was meant to inform. Artificial intelligence (AI) is beginning to close that gap. Over the past several years, a growing body of work has demonstrated that AI models can extract diagnostically relevant molecular and histological information from data sources available in the operating room, in real time, at the point of surgical decision-making. In this review, the authors provide a comprehensive overview of state-of-the-art intraoperative AI models available to neurosurgeons during brain tumor resections. They examine AI-augmented approaches to established intraoperative techniques, such as cryosection and fluorescence-guided surgery, before turning to novel technologies including nanopore sequencing and stimulated Raman histology (SRH). The authors conclude by discussing barriers to widespread clinical adoption and what lies ahead for this field. For cryosection, models such as LGNet and the Cryosection Histopathology Assessment and Review Machine (CHARM) have demonstrated strong performance on intraoperative discrimination of primary CNS lymphoma and molecular subtyping of glioma, respectively, with CHARM achieving areas under the receiver operating characteristic curve of 0.88–0.93 across glioma subtypes. In fluorescence-guided surgery, hyperspectral imaging combined with machine learning has enabled classification of tumor type, grade, and IDH mutation based on the 5-aminolevulinic acid (5-ALA) signal with up to 93% accuracy, extracting molecular information from a modality already present in the operating room. For nanopore-based methylation classification, ensemble models such as Rapid Nanopore Brain Intraoperative Classification (ROBIN) have achieved 90% concordance with the final diagnosis prospectively in under 2 hours. In SRH, foundation models for infiltration detection now outperform FLAIR, contrast enhancement, and 5-ALA fluorescence in identifying tumor margins, while molecular classification models have achieved over 90% glioma subtype accuracy in multi-institutional cohorts. At the single-cell level, morphometric profiling of SRH images has revealed nuclear features distinguishing IDH-mutant from IDH-wildtype tumors, opening new avenues for interpretability. Despite this progress, widespread adoption remains limited by the absence of randomized clinical trials demonstrating patient-level benefit. Addressing this gap and realizing the potential of multimodal AI systems that integrate imaging, pathology, and molecular data represent critical next steps for the field.