by Katrin Berkler, Fraunhofer-Institut für Intelligente Analyse- und Informationssysteme IAIS

Reconstructions of the inferior longitudinal fasciculus (ILF) in the intact hemisphere of a hemispherotomy patient. In the close-up views, corresponding to the area within the red rectangle, streamlines have been clipped near the slice to better visualize the fact that TractSeg reconstructs streamlines within a lesion, whereas the proposed regularized low-rank reconstruction more accurately excludes the lesion, despite segmenting an overall larger tract volume. Credit: NeuroImage: Clinical (2025). DOI: 10.1016/j.nicl.2025.103738

How can nerve pathways in the brain be visualized to improve the planning of complex surgeries? A research team from the Lamarr Institute and the University of Bonn, in collaboration with the Translational Neuroimaging Group at the Departments of Neuroradiology and Epileptology at the University Hospital Bonn (UKB), has investigated an AI-powered method that makes these reconstructions more precise. The study, recently published in NeuroImage: Clinical, could ultimately help make neurosurgical procedures safer.

The brain is a highly complex network of nerve cells interconnected by delicate pathways—known as nerve fibers or tracts. These connections are essential for movement, speech, thought, and many other functions.

To visualize these structures, researchers use tractography, an imaging technique that calculates the course of nerve pathways based on specialized MRI scans. This information is particularly crucial for planning brain surgeries, such as those performed on epilepsy patients undergoing surgical intervention.

Current tractography methods rely on mathematical models that infer the location of nerve pathways from MRI data. However, these methods often involve uncertainties, especially when the brain has been altered due to disease or surgery. This is where modern AI methods come into play: By leveraging machine learning, these systems can recognize patterns and generate more accurate reconstructions.

AI-powered tractography shows potential, but also challenges

In the study, the researchers tested a widely used AI method called TractSeg, originally trained on healthy brains. The team investigated whether it could also work for epilepsy patients who had undergone a hemispherotomy—a surgical procedure that disconnects the two hemispheres of the brain.

The results showed that TractSeg performed well in many cases but also produced unexpected errors: It reconstructed nerve pathways that should no longer exist due to the surgery—a phenomenon known as "hallucination." At the same time, some remaining pathways were either incompletely captured or entirely missing from the reconstruction.

A new hybrid approach for more accurate reconstructions

To address these issues, the team developed a new hybrid method that combines the advantages of AI with the data fidelity of traditional techniques. This approach ensures that only existing nerve connections are reconstructed. The result: No more hallucinations, better detection of preserved pathways, and overall more accurate reconstructions—even in healthy brains.

Prof. Dr. Thomas Schultz, Principal Investigator in the Life Sciences at the Lamarr Institute and a professor at the Institute for Computer Science at the University of Bonn, stated, "Our study demonstrates both the potential and the limitations of AI-powered tractography in clinical applications. Combining AI with traditional methods offers a promising solution for more precise reconstructions, especially when dealing with patient data affected by pathological changes.

"Our goal is to further refine these approaches and make them applicable for neurosurgery in the long run."

More information: Johannes Gruen et al, Deep learning based tractography with TractSeg in patients with hemispherotomy: Evaluation and refinement, NeuroImage: Clinical (2025). DOI: 10.1016/j.nicl.2025.103738

Journal information: NeuroImage 

Provided by Fraunhofer-Institut für Intelligente Analyse- und Informationssysteme IAIS