AI in pituitary surgery: the first real-time clinical test

AI in pituitary surgery has now moved from recorded videos and simulation into a live neurosurgical operating room. During the procedure, a computer-vision system analysed the endoscopic video feed in real time and highlighted important anatomy on a separate screen. The crucial distinction is that the AI did not remove the tumour or decide where to cut: the surgeon remained in control, while the software acted as an additional visual aid.

The case that drew international attention involved Rhys Hibbert, a 48-year-old man in the UK with an approximately 11 mm pituitary tumour that had begun to affect his vision. In May 2026, surgeons at the National Hospital for Neurology and Neurosurgery in London removed the tumour through an endoscopic transsphenoidal approach, entering through the nose. During the operation, an experimental UCL system processed the live video stream and displayed its estimate of key structures that needed to be avoided.

The more informative scientific story, however, is larger than a single patient. The UCL team has released an early clinical series involving six patients with pituitary adenomas. This was a first-in-human feasibility study designed primarily to determine whether the technology could function safely and practically in a real operating theatre.

How the system reads the surgical field

Endoscopic transsphenoidal surgery reaches the pituitary gland through the nasal cavity and sphenoid sinus. The working space is narrow, and the operative field lies close to the carotid arteries, optic nerves and other structures where an error can have serious consequences.

The system uses computer vision and semantic segmentation. Rather than making a surgical decision, it analyses each video frame and marks regions that it predicts correspond to specific anatomical structures. The basic idea resembles object recognition in a phone camera, but the surgical setting is far more demanding: tissue deforms, instruments obstruct the view, blood and lighting alter the image, and a false label can matter clinically.

The model was developed using retrospective surgical video from 98 patients. Researchers selected 1,704 frames from stages where anatomical guidance was considered most useful and annotated structures including the sella, carotid arteries and optic nerves. On a retrospective hold-out dataset, the model achieved a mean intersection-over-union of 76.5% for sella segmentation.

For live use, the analysis ran locally on high-performance hardware rather than relying on cloud processing. That allowed the system to generate overlays with low latency during surgery.

What happened in the first six cases

Six consecutive patients were enrolled. The system was successfully deployed in four operations. In two cases, it was not used at all because a recurring computer reboot bug caused the pre-operative functionality check to fail.

That failure rate is important for interpreting the study. The trial was not designed to prove that AI improves vision, increases the amount of tumour removed or lowers complication rates. Its immediate question was more basic: can a real-time computer-vision system be introduced into an actual neurosurgical workflow without creating unacceptable technical or human-factors problems?

At this stage, the output appeared on a secondary monitor and was treated as an experimental educational and navigational adjunct rather than an autonomous decision-making system.

Among the cases in which the AI was deployed, the investigators reported no AI-attributable clinical complications, no significant distraction of the primary surgeon and no measurable disruption of operating-room workflow. Feasibility ratings were high, with a median score of 4.6 out of 5, while the median System Usability Scale score was 75 out of 100.

The technology was nevertheless far from error-free. Early versions produced false-positive “go-zone” segmentation over carotid regions, and surgical instruments could distort the overlay when they occluded the camera view. Those problems prompted changes to the model, software, hardware and interface between operating sessions.

What the study does not yet show

A six-patient case series cannot establish that the AI reduces complications or improves tumour resection. There was no randomised comparison, no conventional control arm and no formal statistical test of clinical benefit. Two pre-operative technical failures in six attempted deployments also underline how early the platform remains.

The favourable outcome of the first publicly identified patient should not be attributed to the AI itself. Hibbert’s vision improved substantially after surgery, but the study cannot tell us whether the outcome would have been different if the same operation had been performed without the computer-vision overlay.

There is another important limitation: the clinical report is currently a medRxiv preprint and has not yet undergone peer review.

Even so, moving from retrospective video analysis to real-time use during live neurosurgery is a genuine translational step. If larger studies establish reliability and clinical benefit, computer-vision systems could become another layer of surgical navigation — not a replacement for the surgeon, but a way of making critical anatomy harder to lose sight of.