Vision-language models and visual foundation models in neurosurgery

Background Visual foundation models (VFMs) and vision-language models (VLMs) offer adaptable approaches to neurosurgical imaging, segmentation, intraoperative assistance, and clinical decision support. However, the available evidence varies substantially in model type, supervision, validation setting, and clinical task, and their readiness for routine clinical use remains uncertain. Methods We conducted a narrative review of peer-reviewed primary research reports published between January 2020 and August 2026, identified through PubMed, IEEE Xplore, and Google Scholar. Studies evaluating visual, vision-language, or multimodal foundation models in neurosurgical applications were included. Supervised Vision Transformer (ViT) and ViT–CNN systems were retained as contemporary transformer-based comparator studies but were not classified as foundation models. Evidence was synthesised according to model adaptation, clinical application, within-study comparative performance, validation setting, and translational limitations. Results Seven studies met the review criteria. FastGlioma provided a prospective multicentre evaluation of a visual foundation model for intraoperative detection of glioma infiltration. In lumbar-spine MRI, zero-shot SAM and MedSAM performed below supervised nnU-Net on segmentation metrics. A task-adapted SAM2-UNet system achieved a mean Dice score of 0.771 for brain-tumour segmentation, although its mean IoU was 0.569 compared with 0.571 for Mask R-CNN in the reported comparison. Supervised transformer-based systems demonstrated performance in acute traumatic brain injury triage and vertebral fracture classification, although the latter combined ViT, CNN, ensemble fusion, and GAN-based augmentation. In multimodal reasoning, GPT-4o achieved higher absolute diagnostic accuracy than CNS-Obsidian, whereas CNS-Obsidian had higher length-adjusted diagnostic accuracy. Evidence for clinical translation remained limited by heterogeneous validation settings, incomplete reporting of computational performance and safety outcomes, predominantly single-centre or preclinical evaluations, and limited assessment of patient outcomes. Conclusion Foundation-model-based systems show potential across selected neurosurgical tasks, but current evidence is insufficient to establish routine clinical readiness or patient benefit. Reported results are primarily task-specific and should not be interpreted as evidence of a general architectural advantage. Future studies should evaluate these systems prospectively under human supervision while determining whether clinicians can reliably detect and mitigate erroneous outputs. Multi-centre validation, uncertainty assessment, computational efficiency, workflow integration, and patient-centred outcomes should be prioritised.