AI Enters the Operating Room: Real-Time Guidance for Brain Tumour Surgery

Image: operating-room scene associated with brain-tumour surgery coverage

Artificial intelligence has moved from the planning room to the operating theatre in one of the most demanding areas of medicine: surgery at the base of the brain. At University College London Hospitals (UCLH), a clinical trial has reported the first use of an AI system to support a neurosurgeon during a live operation. The system analysed the endoscopic video feed as the procedure unfolded and highlighted critical anatomy near a pituitary tumour, including blood vessels and nerves involved in vision.1 2

The patient, 48-year-old Rhys Hibbert, had an approximately 11-millimetre pituitary tumour that was pressing on the optic nerves and narrowing his peripheral vision. Surgeons removed the tumour through the nose and reported that the procedure protected his sight. Hibbert said his vision improved sharply after the operation and that he was walking independently within a week.1 These are encouraging early results, but they describe one patient in an ongoing feasibility and safety study, not a proven standard of care.

The achievement is important because the system worked from live surgical video rather than relying only on scans taken before the operation. It offered visual guidance at the moment when tissue, instruments and anatomy were changing. The surgeon, not the software, remained responsible for every decision.1

What the London system sees

A pituitary tumour can be difficult to remove safely. The pituitary gland is small and sits close to the carotid arteries and the optic nerves. A narrow error margin separates effective tumour removal from serious complications, including blindness, stroke or major bleeding. Pre-operative magnetic resonance imaging helps the team understand the patient’s anatomy, but it cannot show every detail after instruments enter the surgical field and tissues shift.

During the procedure, an endoscope—a thin camera introduced through the nose—transmitted images to the surgical team. The AI model marked areas where important vessels and nerves were likely to be located and tracked the surgical instruments. The display was intended to help the team remove as much tumour as possible while avoiding structures that might be hidden beneath tissue.2

The model was developed at the UCL Hawkes Institute and trained and evaluated using hundreds of annotated videos from earlier endoscopic pituitary operations. Researchers labelled anatomy, instruments and tissue interactions so that the system could learn patterns that may take an individual surgeon years to encounter in practice.1

Component	Role during surgery	Important limitation
Endoscopic camera	Supplies a continuous view of the surgical field	The view can be blocked by blood, tissue or instruments
AI model	Highlights likely nerves, vessels and other anatomy	A prediction can be wrong or uncertain
Surgeon	Interprets the display and controls the operation	Human expertise remains essential
Clinical trial	Measures feasibility, safety and outcomes	One successful case cannot establish effectiveness

The design reflects a cautious model of medical AI. The software does not make an autonomous incision or command a robot. It provides information, while the clinical team decides whether that information is reliable and useful.

A related approach: identifying tumour cells in seconds

The London operation is not the only way AI is being used around brain-tumour surgery. A separate research programme led by the University of Michigan and the University of California, San Francisco developed FastGlioma, a system that analyses fresh tissue samples rather than a live camera view of the operation.3 4

FastGlioma combines artificial intelligence with stimulated Raman histology, an optical technique that produces microscopic images of unprocessed tissue. In fast mode, an image can be acquired and classified in about 10 seconds. The model estimates whether a sample contains infiltrating tumour tissue that might otherwise look similar to healthy brain.

The study included samples from 220 patients with low- and high-grade diffuse gliomas. The model had been trained using more than 11,000 surgical specimens and four million microscopic fields of view. The full-resolution approach reached an average accuracy of approximately 92%, while the faster mode reached about 90%, according to University of Michigan reporting.4

In the reported comparison, high-risk residual tumour was missed in 3.8% of cases when FastGlioma predictions were used, compared with roughly 24% with conventional methods.3 Those figures are promising, but they should not be interpreted as proof that every patient will benefit. The technology was still investigational and had not received United States Food and Drug Administration approval at the time of the UCSF report.3

The two systems address different moments in the surgical workflow:

AI approach	Input	Output	Potential use
Live endoscopic guidance	Video from the operation	Visual markers for anatomy and instruments	Help avoid critical nerves and vessels during tumour removal
FastGlioma	Microscopic images of fresh tissue	Estimate of tumour infiltration or residual disease	Help decide whether more tissue should be removed or whether additional treatment should be considered

Together, they illustrate a broader shift. AI can support surgeons by interpreting images at a speed that fits the operation, but the value of the system depends on the quality of the images, the training data, the clinical context and the team’s ability to recognise uncertainty.

Why real-time information matters

Brain surgery is a balance between two risks. Leaving tumour behind can allow disease to return, while removing too much normal brain can damage movement, speech, memory or vision. The safest boundary is rarely a simple line visible to the naked eye.

Traditional pathology can provide expert answers, but tissue processing takes time. Intraoperative magnetic resonance imaging can show changes during surgery, but it requires specialised equipment and is not available in every hospital. Fluorescent agents can make some tumour types easier to see, but they do not work equally well for all tumours.4

AI-supported imaging offers another option. A model can compare a new image with a large library of labelled examples and return a probability or visual overlay within the time available to the surgical team. That information may help clinicians decide whether to continue resection, collect another sample or stop because the next step would carry unacceptable risk.

The advantage is not speed alone. A useful system must also communicate what it knows, what it does not know and how confident it is. The UCL team says it studied how confidence should be displayed so that surgeons can distinguish a strong signal from a suggestion that requires caution.1

The safety questions are as important as the technical results

The first live operation raises practical and ethical questions that cannot be answered by an accuracy score. How often does the model fail when the camera angle is unusual? Does performance change across hospitals, surgeons, endoscopes and patient populations? Can a busy operating team understand the display without distraction? What happens when the system loses track of an instrument or marks a vessel in the wrong place?

Clinical trials must also measure patient outcomes, not only image-recognition performance. Relevant outcomes include the completeness of tumour removal, neurological function, complications, recovery time and long-term recurrence. A model that performs well on stored videos may still behave differently in the unpredictable conditions of a live operation.

Governance matters as well. Hospitals will need clear rules for logging AI recommendations, reviewing errors, protecting patient video and assigning responsibility when a clinician disagrees with the system. Training data should represent the diversity of patients and surgical techniques in which the tool will be used. Independent evaluation is necessary before widespread adoption.

From demonstration to dependable clinical tool

The London case shows what becomes possible when an AI system is designed around the operating team rather than placed beside it as a separate research experiment. The system watched the same live field as the surgeons and presented information on a second screen. FastGlioma shows a complementary route: use rapid optical imaging and machine learning to reveal tumour infiltration that may be difficult to identify by sight alone.

Neither approach turns surgery into an automated procedure. Both are decision-support tools. Their promise lies in giving specialists additional, timely information while preserving human judgement and accountability.

The next stage will require larger studies, transparent reporting and careful comparison with existing methods. If those studies confirm that AI can improve the balance between tumour removal and protection of healthy brain tissue, real-time assistance may become a valuable part of precision neurosurgery. For now, the most accurate description is more measured: AI has entered the operating room as an experimental assistant, and early clinical work is testing whether that extra pair of eyes can make delicate surgery safer.

References

[1] First patient in live AI assisted sight-saving brain surgery — UCLH

[2] World’s first patient to undergo live AI-assisted brain surgery has tumour removed — BBC News

[3] This AI Tool Helps Neurosurgeons Find Sneaky Cancer Cells — UCSF

[4] In 10 seconds, AI model detects cancerous brain tumor often missed during surgery — Michigan Medicine

Editorial note: This is a reported science article, not medical advice. The clinical trial described by UCLH is ongoing, and investigational tools should not be presented as routine treatment.