In the days leading up to a neurosurgical procedure, clinicians use high-resolution MRI and CT scans to build a meticulous 3D map of the patient’s brain. This map identifies the exact coordinates of a tumor or the precise "highway" of a white matter tract. However, the moment the surgeon opens the skull, this map begins to expire.
The brain is not a rigid structure; it is a soft, deformable organ. Once the cranium is opened, changes in intracranial pressure, the drainage of cerebrospinal fluid, the effects of gravity, and the physical removal of tissue cause the brain to sag and deform. This phenomenon, known as brain shift, creates a spatial discrepancy between the static preoperative images and the dynamic intraoperative reality.
Research in this field focuses on closing this "reality gap" by developing methods to track, predict, and compensate for these deformations in real-time.
The Failure of Static Navigation
Standard neuronavigation systems act like a GPS for the brain, but they often rely on "offline" maps. As H Iseki and colleagues noted in their work on Intelligent Operating Theaters, these maps become increasingly inaccurate as surgery progresses.
This inaccuracy has severe consequences for modern surgical tools:
- Augmented Reality (AR): When digital 3D models are overlaid on a surgeon’s view, brain shift causes a "navigation mismatch." S Fujimoto documented cases where targets shifted by as much as 3 cm, rendering AR overlays misleading.
- Robotics: Surgical robots require sub-millimeter precision. Zhuojin Wu argues that without real-time feedback to account for tissue displacement, robotic guidance loses its primary advantage.
- Neuromodulation: Placing deep brain stimulation (DBS) electrodes or spinal cord arrays requires hitting tiny neural targets. Jaimie M. Henderson and Samuel R Parker highlight that even minor physical migrations can move an electrode away from the intended circuit, necessitating invasive revisions.
Real-Time Sensing: Seeing the Shift
To correct the map, surgeons need a way to "see" the brain’s current state without stopping the surgery for a full MRI.
One robust solution is Intraoperative Ultrasound (iUS). Unlike MRI, ultrasound is fast and portable. Ingerid Reinertsen and Maxime Descoteaux demonstrated that by tracking blood vessels using Doppler ultrasound, algorithms can register the current position of the brain to preoperative scans in under a minute. Similarly, Kecheng Shen showed that microvascular Doppler can act as a real-time "GPS" for critical arteries that may have shifted during skull base surgery.
However, ultrasound images look vastly different from MRI scans—a problem known as the "appearance gap." To bridge this, researchers like Daniil Morozov and Reuben Dorent have developed CrossKEY, which uses patient-specific synthetic images to help AI recognize landmarks across different imaging types.
Predicting Deformation with Physics and AI
If we know the physical properties of brain tissue, can we predict how it will shift before it happens?
Traditionally, this was done using the Finite Element Method (FEM), which simulates the brain as a complex mesh of physical particles. While accurate, FEM is computationally "heavy" and often too slow for the operating room.
The new frontier involves Physics-Driven AI. Yasmin Salehi and Dennis Giannacopoulos introduced PhysGNN, a Graph Neural Network that learns the laws of physics. It achieves the accuracy of traditional simulations but operates thousands of times faster. By treating the brain as a graph, the model can ingest sparse data from the surgical site and instantly "fill in the blanks" to show how deep structures have moved.
Signal-Based Localization
When visual maps fail, researchers turn to the brain’s own electrical signals. This is particularly useful for placing electrodes in regions where anatomical landmarks are distorted.
Akshay Rao and Kelvin L. Chou have shown that by correlating intraoperative microelectrode recordings with preoperative MRI, targeting errors can be reduced significantly. Taking this further, Tianxiao He developed Lfp2vec, a transformer-based model that can identify a probe's location simply by "listening" to the raw local field potentials of the surrounding neurons. This "functional GPS" bypasses the need for physical maps entirely, relying instead on the unique electrical signature of different brain regions.
The Path to Dynamic Navigation
The shift from static to dynamic navigation represents a fundamental change in how we approach the brain. By combining large-scale datasets like ReMIND (the Brain Resection Multimodal Imaging Database) with real-time sensing and physics-informed AI, the field is moving toward a future where the surgical map evolves alongside the patient.
As Reuben Dorent emphasizes in the ReMIND2Reg Challenge, the goal is to provide surgeons with accurate guidance during the most critical phases of surgery—post-resection—when the brain has deformed the most and the distinction between tumor and healthy tissue is hardest to see.