What enables an experienced neurosurgeon to spot an abnormality on an MRI scan almost immediately, whilst a medical student takes much longer to do so? According to researchers at Macquarie University in Australia, the answer lies not only in medical knowledge, but also in the way experts process visual information. Two new studies show that the eye movements of experienced specialists exhibit clear, measurable patterns. The researchers believe that these insights could contribute to the development of artificial intelligence (AI) that not only analyses medical images but also learns to ‘see’ in the same way as an expert does.
According to the researchers, the results, published in *Medical & Biological Engineering & Computing* and the *Journal of Eye Movement Research*, offer a new approach to both AI development and the training of future doctors. The research was carried out by Macquarie University’s Computational NeuroSurgery Lab (CNS Lab).
Visual expertise proves to be measurable
The team investigated how the eye movements of medical students, trainee doctors and experienced neurosurgeons differ when assessing medical images, ranging from X-rays to CT and MRI scans. To this end, eye movements were recorded with high precision. The researchers then analysed this data using fractal mathematics and machine learning. Fractal analysis makes it possible not only to record where someone is looking and for how long, but also how organised – or, conversely, how random – the overall viewing pattern is.
According to the researchers, the way people view medical images changes fundamentally as their expertise increases. Whilst beginners primarily respond to striking visual features, experts develop a more structured way of scanning, directing their attention more quickly and specifically towards diagnostically relevant areas.
From student to specialist
In the first study, the researchers followed thirteen medical students over three consecutive semesters. The participants viewed various medical images whilst their eye movements were recorded. The analyses were consolidated into a new metric, the Fractal Eye-Gaze Expertise Index (FEI). This index revealed a clear progression: as students progressed through their training, their eye movements became less erratic and more similar to those of experienced specialists.
According to the researchers, this development reflects a shift from general image perception to goal-oriented diagnostic interpretation. Thanks to fractal analysis, this growing visual expertise could be objectively quantified for the first time. The researchers emphasise that expertise is therefore not expressed solely through greater knowledge, but also through a fundamentally different way of looking at things.
AI recognises level of experience
A second study expanded the research to a larger group of 69 participants, comprising people with no medical experience, doctors training to become neurosurgeons, and experienced neurosurgeons. They were asked to assess both normal and abnormal MRI scans of the brain.
This revealed that experienced specialists navigate the images more efficiently, spend more time focusing on abnormal areas and adapt their viewing strategy to the type of pathology they are assessing.
The researchers then developed an AI model that analysed both the duration of gaze fixations and the three-dimensional fractal characteristics of the viewing pattern. This enabled the system to distinguish between novices, trainee doctors and experts with an accuracy of over 93 per cent. According to the study, this demonstrates that visual expertise translates into a recognisable pattern that can be learnt and classified by AI.
A new building block for medical AI
The researchers view the results as more than just a study of eye movements. In their view, the patterns exhibited by experts during image analysis reflect complex cognitive processes. By using these patterns as training data, AI would not only learn which abnormalities are visible, but also how experienced doctors distribute their attention across an image. This could be valuable for future AI systems that analyse radiological images or support doctors in diagnostic decision-making. In addition, the method offers opportunities to objectively monitor the development of visual expertise during medical training and possibly even provide more targeted guidance.
At the same time, the researchers emphasise that the aim is not to replace doctors, but to gain a better understanding of human expertise. According to them, the central question is not only how machines can learn, but above all which human skills we want to teach AI. The fractal patterns in the eye movements of experts show that experience determines not only what someone sees, but also how someone looks. It is precisely this way of looking that could serve as an important source of inspiration for the next generation of medical AI systems in the future.
“We believe that our findings have important implications for the way in which trainee doctors are assessed, how training programmes for radiology and neurosurgery are structured, and how AI systems can ultimately be trained to interpret medical images in the same way as an expert does,” said Professor Di Ieva, who founded the CNS Lab in 2018.
AI in surgery
Earlier this year, an AI co-pilot was unveiled in Ghent, Belgium, designed to support surgeons in real time during robotic operations. The system analyses live surgical images, recognises surgical phases and anatomical structures, and provides contextual support for complex decisions. The technology builds on an earlier innovation from 2023, in which four AI models were deployed simultaneously during a robot-assisted kidney operation. The new AI co-pilot supports the entire surgical process, from pre-operative planning to intra-operative guidance and training.
For further development, Orsi is using an NVIDIA supercomputer that provides the necessary real-time computing power. The technology was tested live during a robotic operation at AZORG Hospital. According to Orsi, this development marks an important step towards the wider clinical application of AI in surgery, although validation, regulation and integration into healthcare remain key areas of focus. Another development is the use of AI to train trainee surgeons.
References
Medical & Biological Engineering & Computing (Research)
Journal of Eye Movement Research (Research)
