For generations, the human brain has been the ultimate frontier – a universe of thought and memory, protected by a thick skull and wrapped in mystery. Operating on the brain is like navigating a dense, shifting fog: one wrong move can change a life forever. Unfortunately, surgeons often need to do so under the direst circumstances.
Once cancer cells invade the brain, the prognosis is often severe, and surgery is the primary treatment for brain tumours. The most common brain or spinal tumor is glioma, and over 90% of patients progress to advanced stages within five years. To enable surgeons to remove tumors precisely, high-resolution images of the affected brain regions are essential. However, the brain is encased by a thick skull and has a complex structure, making neurosurgery a major clinical challenge – operating inside the brain is like navigating in dense fog, where a slight mistake can permanently alter a patient’s life. Recent advances in cone-beam computed tomography (CBCT) have brought breakthroughs. By integrating preoperative and intraoperative imaging, CBCT can provide dynamic, real-time views of the brain, offering strong support for surgical decision-making and precision treatment.
“This enables real-time tracking and precise localisation of lesion areas during surgery, significantly enhancing accuracy and safety while reducing damage to healthy brain tissue,” says Professor Yuan Yixuan, Associate Professor in the Department of Electronic Engineering at CUHK. “CBCT offers several advantages over other imaging modalities, including a low radiation dose, high spatial resolution, compact and lower-cost equipment with a small footprint suitable for clinic settings, fast acquisition and reconstruction, flexible patient positioning and usefulness for intra-procedural image guidance and verification.”
When images don’t line up: the hidden flaws in brain scans
However, the technology has been around for less than 30 years, and there are still important challenges to overcome. Imaging devices can be limited and inflexible. Due to the mismatch between preoperative and intraoperative medical data, it is useful to combine different brain modality data, such as CT and MRI, which produces more accurate results – unfortunately, though, scanning devices differ and the people being scanned are often positioned slightly differently, meaning the results often aren’t aligned with each other. And there’s a lack of information about the brain’s fibre tracts, which help different bits of the brain talk to each other. Lack of such information may cause damage to critical neural pathways during surgeries.
From curiosity to neurosurgical innovation
Professor Yuan’s journey into brain imaging began not in medicine, but in engineering. As an undergraduate, she focuses on precision medicine and aims to developing AI model to automatic medical data analysis and decision making. After a PhD at CUHK and a postdoctoral fellowship at Stanford, she returned to Hong Kong: to build better tools for seeing the unseen medical world.
Her most important previous project was the dynamically named NeuroSTORM (it stands for Neuroimaging Foundation Model with Spatial-Temporal Optimised and Representation Modelling). Functional Magnetic Resonance Imaging (fMRI) is crucial for studying brain function and diagnosing neurological disorders. However, existing methods for fMRI analysis suffer from reproducibility and transferability challenges, due to complex pre-processing pipelines and task-specific model designs. NeuroSTORM represents the first efforts in fMRI foundation model. It learns generalisable representations directly from 4D fMRI volumes, and achieves efficient pre-trained knowledge transfer across diverse downstream applications. Her latest work aims to promote the CBCT based brain data analysis in the neurosurgery and improve the accuracy.
A joint four-year project with Southeast University, Nanjing, it will consist of a mobile dual-energy CBCT imaging system; high-quality CBCT imaging algorithms; technology to combine images from a range of technological sources; and immersive augmented surgical assistance with diffusion tensor imaging, which can provide images of the brain’s white matter (tracts). In this collaboration work, Southeast University is focusing on the hardware design, while the CUHK team is developing the AI algorithms and software.
“We are building a unified neurosurgery model that can synthesise missing imaging modalities and integrating augmented reality into neurosurgical simulation to create immersive, interactive surgery environments,” she says. “This project aims to develop a comprehensive neurosurgery navigation system, integrating the hardware and software design and addressing key challenges in neurosurgical navigation, including intraoperative real-time imaging, soft tissue visualisation, multi-modal image registration and brain function preservation.”
When machines assist surgeons: the future, ethics and safety
Professor Yuan’s team is building this model using artificial intelligence (AI), and the advantages are clear. Current AI advances are rapid: virtual cells can now mimic lab experiments, which speed up testing and reduce the need for physical trials. AI agents can automate complex tasks like designing experiments and organising data, while generative tools help us create high-quality images and summaries to share our work more effectively.
“AI is transforming the way we do research, but it comes with challenges. Before we use these tools in clinics or make new policies, we must strictly verify the results to ensure they are reliable. We also need strong safety standards and ethical oversight to protect patients and the public.”
The use of AI in medical image analysis presents further ethical issues where the security and privacy of data is concerned, she adds. It also poses questions of who is responsible if something goes wrong. But if these issues can be overcome, the AI technology promises to unlock a future in which neurosurgeons can perform their miracles with ever increasing accuracy and safety.
