AI outlined anatomy during six pituitary operations in an eight-patient London series, published on 3 October 2026. A computer startup fault prevented the other two deployments.[1]
The surgeon operated using the usual main screen. AI outlines appeared on a second display, activated by foot pedal. The system initially supported education and orientation, rather than primary surgical decisions.[1]
UCL reported in August that the first participant’s tumour was removed and his vision improved. The small clinical series establishes feasibility; it does not show that AI caused the improved outcome.[2][1]
The display is part of the medical tool
Pituitary surgery takes an endoscope through the nose towards a crowded region at the skull base. Surgeons must judge an opening large enough to reach the tumour while protecting neighbouring structures. A 2024 UCL design study explored how extra visual information could support that judgement.[3]
What surgeons wanted from an overlay
| Design choice | Purpose |
|---|---|
| Visible anatomical boundaries | Help judge the extent of the opening.Source 3 |
| Avoid solid visual blocks | Keep the actual tissue visible.Source 3 |
| Optional highlighting | Let the surgeon choose which structures to inspect.Source 3 |
Four surgeons evaluated the interactive prototype in the second part of that study. They differed over the best rendering style, but valued seeing the spatial relationship between structures. An informative outline has to remain readable without concealing the scene it is meant to explain.[3]
A live image keeps changing
A separate 2025 study, ConsisTNet, investigated a difficulty peculiar to video: a prediction can jump between frames even when the anatomical scene has barely changed. Its model combined information from three consecutive frames and used tracking to extend sparse human annotations across the sequence.[4]
The system produced both segmented regions and landmark positions. These answer different questions: a region outlines an area, while a landmark identifies a location. Training also penalised inconsistent predictions across neighbouring frames, encouraging the display to follow anatomical continuity.[4]
ConsisTNet’s improvements were measured on recorded surgical video. The live clinical trial used a separate model.[4][1]
What the first operating-theatre cases exposed
During early cases, the model sometimes marked the sella—the bony compartment housing the pituitary—as extending over carotid-artery regions. Instruments blocking the camera could also distort its outlines. An apparent target area could therefore cover anatomy that surgeons need to protect.[1]
The team added training examples and refined the model and firmware. Displayed confidence describes a model prediction, not a measured probability of surgical safety. The overlays remained on the separate, optional screen.[1]
The authors reported no AI-related clinical complications in this small series. Planned larger single-centre and multicentre studies will examine performance beyond the six successful deployments; an effect on clinical outcomes remains to be measured.[1]
Sources & further reading
- First-in-human evaluation of real-time pixel-level AI-assisted anatomical segmentation in neurosurgery: pituitary surgery as an exemplar Pituitary · 3 Oct 2026 · Accessed 12 Oct 2026
- First patient in live AI-assisted sight-saving brain surgery UCL · 26 Aug 2026 · Accessed 12 Oct 2026
- Augmented reality for endoscopic transsphenoidal surgery: evaluating design factors with neurosurgeons International Journal of Computer Assisted Radiology and Surgery · 26 Jul 2024 · Accessed 12 Oct 2026
- ConsisTNet: a spatio-temporal approach for consistent anatomical localization in endoscopic pituitary surgery International Journal of Computer Assisted Radiology and Surgery · 29 Apr 2025 · Accessed 12 Oct 2026