Image-guided neurosurgery relies on a fundamental assumption: the brain remains stationary between the time a preoperative scan is taken and the time the surgeon operates. In reality, this assumption breaks down the moment the surgical procedure begins.
When a surgeon performs a craniotomy, the physical state of the brain changes dramatically. Cerebrospinal fluid (CSF) drains out, intracranial pressure drops, and gravity pulls the soft tissue downward. Furthermore, as tumors are resected, the surrounding brain tissue collapses into the newly created voids. This physical displacement and deformation of tissue is known as intraoperative brain shift.
Because of brain shift, high-resolution preoperative maps—such as 3D structural MRIs or diffusion tensor imaging (DTI) tractographies—lose spatial accuracy as the operation progresses. This creates a critical challenge: navigating near delicate functional boundaries using a static, outdated map can lead to incomplete tumor removals or accidental damage to vital pathways.
To solve this problem, researchers are developing methods to dynamically update surgical maps, model tissue deformation in real time, and alter surgical workflows to mitigate tissue movement.
The Mismatch Between Static Maps and Dynamic Anatomy
The consequences of brain shift are particularly severe when operating near eloquent areas of the brain responsible for movement, sensation, or language. As Takahiro Matsuhashi et al. demonstrate in their work on Identification of IFOF during Insular Glioma Surgery, preoperative 3D fiber tracking maps of critical language highways, like the Inferior Fronto-Occipital Fasciculus (IFOF), lose accuracy during tumor removal. Relying solely on the initial scan can cause surgeons to misestimate the true location of these tracts.
Similarly, in deep brain stimulation (DBS) surgeries, sub-millimeter precision is required to place electrodes into tiny gray matter nuclei. Yuval Duchin et al. note in Patient-specific anatomical model for deep brain stimulation based on 7 Tesla MRI that even a slight shift after opening the skull introduces a mismatch between high-field 7T preoperative scans and the actual intraoperative positions of internal structures. As Jaimie M. Henderson argues in “Connectomic surgery”, this physical displacement forces surgeons to constantly account for dynamic changes in brain geometry rather than trusting static anatomical scans.
[Preoperative Scan] ──(Craniotomy & CSF Leak)──> [Brain Shifts/Sags] ──> [Mismatch with Navigation Map]
Real-Time Intraoperative Imaging
The most direct way to correct for brain shift is to acquire new images during the surgery and warp the preoperative maps to match the new reality.
One approach is intraoperative MRI (iMRI). Luoyi Tian et al. show in Clinical evaluation of resection of functional area gliomas guided by intraoperative 3.0 T MRI combined with functional MRI navigation that updating navigation maps with real-time iMRI data allows surgeons to recalibrate their spatial coordinates. This intraoperative update compensates for tissue displacement caused by CSF loss and tumor mass removal, leading to higher rates of complete tumor resection. This builds on early architectural frameworks like the "Intelligent Operating Theater" described by H Iseki et al. in Intelligent Operating Theater in Neurosurgical Field, which integrated iMRI scanners directly into the surgical workflow.
However, iMRI is expensive, time-consuming, and interrupts the flow of surgery. Intraoperative ultrasound (iUS) offers a faster, more dynamic alternative. To bridge the gap between preoperative MRI and intraoperative ultrasound, researchers use non-rigid registration algorithms.
Ingerid Reinertsen et al. developed a solution detailed in Vessel Driven Correction of Brain Shift that uses intraoperative Doppler ultrasound to map the brain's blood vessels. Because blood vessels form a distinct, continuous network embedded within the brain tissue, tracking how these vessels move provides a mathematical blueprint of how the surrounding tissue has deformed. By aligning the intraoperative ultrasound vessel maps with preoperative MRI scans, their algorithm corrects for brain shift in under a minute.
To make these ultrasound updates more robust, Jie Luo et al. introduced A Feature-Driven Active Framework for Ultrasound-Based Brain Shift Compensation, which combines feature matching with Gaussian Processes to actively calculate and compensate for non-rigid tissue deformations. To train the machine learning models underpinning these multi-modal alignments, large-scale open datasets like the one presented by Parikshit Juvekar et al. in ReMIND: The Brain Resection Multimodal Imaging Database provide paired preoperative MRI, iMRI, and iUS data to benchmark deformation recovery algorithms.
Biomechanical Modeling and Machine Learning
When continuous intraoperative imaging is unavailable, surgeons can use computational models to predict how the brain will deform.
Traditionally, this is done using physics-based simulations like the Finite Element Method (FEM) or Linear Elastic Models (LEM). For instance, in A Comprehensive System for Intraoperative 3D Brain Deformation Recovery, the authors utilized a linear elastic model to simulate volumetric deformation based on surface measurements. Similarly, Kun Wu et al. demonstrated in Nonrigid 3D Brain Registration Using Intensity/Feature Information that intraoperative stereo cameras can track the exposed surface of the brain, using a Bayesian framework to propagate those surface shifts down into the deeper structures of the volumetric preoperative scan.
The primary limitation of traditional biomechanical models is computational latency; solving complex differential equations for soft tissue physics takes too long for live surgical navigation.
To overcome this speed bottleneck, Yasmin Salehi et al. developed PhysGNN, detailed in PhysGNN: A Physics-Driven Graph Neural Network Based Model for Predicting Soft Tissue Deformation in Image-Guided Neurosurgery. PhysGNN treats the surgical mesh of the brain as a graph, where nodes represent tissue points and edges represent physical connections. By training a Graph Neural Network (GNN) on boundary conditions and physics equations, the model learns to predict complex, non-rigid soft tissue deformations thousands of times faster than traditional FEM simulations, achieving near-instantaneous updates during live navigation.
[Surface Shift Captured by Camera/US] ──> [Graph Neural Network (PhysGNN)] ──> [Instantaneous Volumetric Prediction]
Mitigating Shift Through Procedural Design
While imaging and modeling focus on correcting brain shift after it occurs, some research directions focus on preventing it through smart procedural design.
In surgeries that require inserting deep probes into the brain, such as stereoelectroencephalography (SEEG) for epilepsy monitoring, the order of surgical steps matters immensely. Ayataka Fujimoto et al. address this in Neuronavigation-guided Frameless Stereoelectroencephalography (SEEG).
If a surgeon opens a wide area of the dura (the brain's protective membrane) to place surface grids first, CSF escapes rapidly, causing the brain to sag and shifting deep targets away from their preoperative coordinates. Ayataka Fujimoto et al. demonstrate that by sequencing the surgery to place deep depth electrodes through a closed dura before exposing the surface, the intracranial environment remains sealed. This prevents fluid loss and subsequent tissue displacement, ensuring that deep targets remain exactly where the preoperative imaging predicted.
Key Foundational Work
Use of Propofol for Awake Craniotomy in Epilepsy Surgery (1997) — contributed 1 concept elements to this direction.
Surgical and Functional Anatomy for Presurgical Planning : The Frontal Language Areas (2001) — contributed 1 concept elements to this direction.