New Automated Tractography Tool Maps Brain Tumour Pathways in Minutes Instead of Hours

Neurosurgeons planning operations on children with brain tumours face one of the most delicate mapping problems in medicine: identifying the bundles of white matter that carry movement, language and vision signals, and steering clear of them. A new study published in Frontiers in Neuroimaging introduces a tool that promises to make that mapping dramatically faster and easier, converting an automated tract-segmentation method into the streamline format that clinicians and researchers actually use, without sacrificing the speed that made it attractive in the first place.

The technique, called Tractfinder-constrained Tractography, or TcT, was developed by researchers at UCL Great Ormond Street Institute of Child Health, the Francis Crick Institute and Great Ormond Street Hospital for Children in London. It builds on Tractfinder, an automated method introduced in recent years that outlines white matter tracts in patients whose brain anatomy has been distorted by tumours. Tractfinder’s strength lies in its speed and automation, but its outputs arrive as voxel-based masks, essentially three-dimensional volumes marking where a tract is likely to sit. That format creates a practical bottleneck, because most established tractography tools, visualisation workflows and downstream analysis methods are built around streamlines, the continuous fibre pathways that model tracts as collections of curved lines through the brain.

The distinction matters more than it might sound. A binary mask tells a surgeon where a tract lies but offers little of the geometric detail that streamline representations provide. Streamlines also enable analyses that voxel outputs cannot support, such as along-tract statistical profiling, which examines how tissue properties change along the length of a fibre bundle. By constraining probabilistic tractography to the probability maps that Tractfinder produces, TcT generates streamline representations while preserving the automation that makes the original method clinically appealing. The Tractfinder-derived segmentations serve as binary mask regions of interest, restricting streamline propagation to within the tract volume and ensuring that the reconstructed streamlines remain consistent with the underlying segmentation.

To test the approach, the team applied TcT to ten pre-operative paediatric patients with supratentorial tumours, a group whose diagnoses ranged from pleomorphic xanthoastrocytoma and craniopharyngioma to ependymoma, pilocytic astrocytoma, epidermoid cyst, glioneuronal tumour and low-grade glioma, with some tumours affecting both hemispheres. Three clinically critical tracts were selected for evaluation: the corticospinal tract, which carries motor signals; the arcuate fasciculus, a key language pathway; and the optic radiation, which relays visual information. These pathways are frequently displaced, infiltrated or disrupted by supratentorial tumours, and surgical injury to any of them carries significant functional consequences.

The imaging data came from routine pre-operative assessments performed on a Siemens MAGNETOM Prisma 3 Tesla scanner. Structural T1-weighted images were acquired at isotropic 1 mm resolution, while diffusion-weighted images used a multi-shell scheme with 60 directions at b-values of 1,000 and 2,200 s/mm², plus 13 unweighted volumes, at 2 mm isotropic resolution. After standard preprocessing steps including thermal noise removal, Gibbs ringing correction, motion and distortion correction, and bias field correction, the diffusion data were linearly registered to MNI152 standard space. The researchers deliberately chose affine rather than non-linear registration, reasoning that a global transformation accounts for individual anatomical variability while limiting the streamline distortions that local registration errors can introduce, and that affine methods offer advantages in speed and reproducibility.

Fibre orientation distributions were estimated using multi-shell multi-tissue constrained spherical deconvolution, with tissue response functions derived through an unsupervised algorithm. Where tumour masks were needed, the team employed an nnU-Net-based segmentation framework validated for routinely acquired, heterogeneous clinical data. Critically, that framework was trained on incomplete imaging datasets, so lesion masks could be derived from T1-weighted images alone, reducing both data requirements and processing time. Tractfinder then generated pseudo-probability maps of tract location, computed as voxel-wise inner products of a registered orientation atlas and the subject’s fibre orientation images. Where tumour mass effect had displaced a tract, a tumour deformation model incorporating the lesion mask was applied. Manual adjustment of the deformation parameter was required in only one of the ten cases, suggesting the pipeline is largely robust to automation, though the authors note that visual quality control remains advisable.

The probability maps were thresholded at 0.05, a value determined empirically in the original Tractfinder methodology, and the resulting binary masks were used to constrain targeted probabilistic streamline tractography using the iFOD2 algorithm. Tractography ran with high-density seeding of one million seeds and no fixed streamline count, with MRtrix3 default settings governing step size, angle threshold and fibre orientation cutoff. For comparison, the same datasets underwent conventional targeted probabilistic tractography seeded from manually placed regions of interest, following a published protocol with specific seed, inclusion and exclusion regions for each tract. Conventional streamlines were then post-processed to remove spurious fibres using a track-density-imaging filter that rejected any streamline not passing through voxels meeting a minimum density threshold.

The results were striking. Spatial agreement between the two approaches, quantified with the Bundle Adjacency metric, ranged from 2.1 to 2.6 mm on average across the three tracts, with the optic radiation showing the closest correspondence at 2.10 mm and the arcuate fasciculus the widest at 2.57 mm. Those figures compare favourably against published benchmarks for conventional probabilistic tractography, where mean Bundle Adjacency between different pipelines has been reported at 4.3 mm, and they approach the roughly 2 mm variability seen when the same protocol is repeated. In other words, the automated pipeline agreed with the manual one about as closely as expert users agree with themselves. The agreement held across a clinically heterogeneous cohort encompassing a range of tumour types, locations and degrees of tract displacement, lending confidence to the approach’s generalisability.

The practical gains were equally significant. The TcT pipeline ran as a single automated script on a standard MacBook Air laptop, requiring no dedicated GPU infrastructure, and completed in approximately 5 to 15 minutes per subject, excluding preprocessing. The automated tumour segmentation step added only about 10 to 15 seconds per patient on GPU-accelerated hardware. Conventional tractography, by contrast, took one to two hours per subject and demanded manual placement of seed, inclusion and exclusion regions for every tract, a process that requires considerable neuroanatomical expertise and introduces inter-rater variability. The authors suggest this speed advantage could prove particularly valuable in time-critical scenarios such as intraoperative MRI, where rapid tract delineation is essential for surgical decision-making when tumours sit close to eloquent white matter pathways.

The researchers are careful to frame the findings as agreement rather than validation. Because ground truth tract anatomy is unavailable in this population, TcT was benchmarked against conventional tractography rather than against an independent standard, and the observed agreement may partly reflect shared anatomical assumptions embedded in the atlas-derived priors that constrain both the segmentation and the tractography. The cohort was also small and dominated by low-grade tumours, which limits generalisability to infiltrative gliomas and high-grade lesions associated with extensive peritumoral oedema, where automated delineation may face greater challenges. Bundle Adjacency, moreover, captures spatial proximity but not topological correspondence. Still, the authors argue that TcT represents a meaningful step toward scalable tract delineation, and future work will evaluate the method in larger multicentre cohorts, extend it to additional clinically relevant tracts, formally assess reproducibility across raters and sites, and explore its integration into prospective surgical planning workflows. If those efforts succeed, the minutes-long mapping of a child’s motor, language and visual highways could become a routine part of the operating theatre’s toolkit.

Subject of Research: Automated streamline-based white matter tract segmentation for pre-operative paediatric brain tumour imaging

Article Title: A rapid streamline-based extension of Tractfinder for white matter tract segmentation

Article References: Kanel, D., Young, F., Seunarine, K. K., Clark, C. A., Aquilina, K., & Clayden, J. D. (2026). A rapid streamline-based extension of Tractfinder for white matter tract segmentation. Frontiers in Neuroimaging, 5, Article 1873040. https://doi.org/10.3389/fnimg.2026.1873040

Image Credits: AI Generated

DOI: 10.3389/fnimg.2026.1873040

Keywords: tractography, white matter, diffusion MRI, paediatric brain tumours, Tractfinder, surgical planning, corticospinal tract, arcuate fasciculus, optic radiation, bundle adjacency, neuroimaging, automation