Medical Technology · uk
An Extra Pair of Eyes for Pituitary Surgery: UCL Tests AI to Identify Key Structures in Real Time
AI trained on hundreds of surgical videos has entered the operating room to help surgeons identify important tissues. No AI-related complications were reported in the first six cases, but comparative evidence on whether it can better protect vision and hormonal function is still lacking.
Removing a pituitary tumour tests more than the ability to locate the tumour; it also requires distinguishing nearby blood vessels and optic nerves. University College London (UCL) is bringing artificial intelligence into this confined surgical space: analysing live images from an endoscope and marking key anatomical structures to provide surgeons with additional visual cues. The system has begun evaluation in humans. The evidence so far primarily addresses whether it can be used safely and practically, and is not yet sufficient to demonstrate that it improves surgical outcomes.
On November 5, UCL presented the research, a collaboration between Sophia Bano and neurosurgeon Hani Marcus. The pituitary gland sits at the base of the skull, and surgery can be performed through the nasal cavity using an endoscope. However, tumours may be only a few millimetres from important blood vessels and nerves, while blood and other bodily fluids can obscure the view. The team trained the model on hundreds of surgical videos annotated by experts, enabling the system to learn to identify anatomical structures and differences between patients, then overlay its findings on the images.
How these markings are displayed also matters for surgical judgment. An explanation published in September by the UK's Pituitary Foundation stated that the AI view is displayed separately from the standard surgical images, is switched off by default, and is activated when the surgeon presses a foot pedal. It also shows the model's confidence in its predictions. Surgeons retain control over the procedure and decisions, while the confidence information allows them to assess the system's uncertainty when consulting the markings.
The system has entered clinical research at the National Hospital for Neurology and Neurosurgery in London. In August, UCL published an account of the first patient's surgery, reporting that the patient's vision was preserved after tumour removal. This provides an example of use in practice, but a good outcome in a single patient cannot establish how much additional help the AI provided, nor can it be used to infer outcomes for other patients.
The Foundation article, co-authored by Marcus, further outlined the limits of the preliminary study: in the first six operations in which the system was successfully used, no AI-related complications were reported, and no obvious disruption to the surgeon or surgical team was observed. However, the study was small and was not designed to demonstrate a reduction in complications. Patients' hormonal outcomes and lengths of hospital stay were broadly similar to those of previous patients who underwent conventional surgery. The next step is to involve more surgical teams, before expanding the evaluation to multiple hospitals.
The November presentation also covered another application: researchers used machine learning to analyse digital slides of brain tissue donated after death, quantifying abnormal proteins associated with neurodegenerative diseases and examining their amounts and distribution. These tools are used to study differences between diseases and represent a separate application from real-time surgical guidance. The presentation did not provide new controlled clinical results.
For surgical teams, the next key question is whether the identification markings can provide reliable information when it is actually needed. Further research must establish whether the initial practical feasibility can translate into fewer blood vessel or nerve injuries, and better preservation of vision and pituitary function. Until that evidence is available, the clinical value of this AI should continue to be understood in terms of an assistive tool under evaluation.