Source-linked AI summary
TractSeg - Fast and accurate white matter tract segmentation
Jakob Wasserthal, Peter Neher, Klaus H. Maier-Hein
TL;DR
White matter tract segmentation is important for analyzing healthy and diseased brains, but existing approaches can be time-consuming, subjective, and computationally intensive. TractSeg directly segments tracts from fODF peaks with a convolutional neural network, avoiding several conventional pipeline steps; across 105 HCP subjects, it achieved unprecedented accuracy and was less affected by reduced-resolution clinical-quality data.
Problem
Accurate white matter tract segmentation is needed for tractometry, while common virtual dissection and multi-step automated pipelines are time-consuming, subjective, or prone to accumulated inaccuracies.
Method
TractSeg uses encoder-decoder fully convolutional neural networks to directly produce multi-label tract segmentations from fODF peak fields without registration, parcellation, tractography, or clustering.
Results
Across 72 tracts in 105 HCP subjects, TractSeg achieved unprecedented accuracy and was less affected by reduced resolution in clinical-quality data than the benchmark methods.
Takeaways & Limitations
Direct fODF-peak segmentation provides a faster, easier-to-set-up alternative to pipelines requiring tractography, registration, parcellation, or clustering.
Takeaways & Limitations
TractSeg depends on the availability and quality of supervised training data, and domain shifts from different scanners or acquisition sequences can reduce performance.
Abstract
from arXiv · showhide
The individual course of white matter fiber tracts is an important key for analysis of white matter characteristics in healthy and diseased brains. Uniquely, diffusion-weighted MRI tractography in combination with region-based or clustering-based selection of streamlines allows for the in-vivo delineation and analysis of anatomically well known tracts. This, however, currently requires complex, computationally intensive and tedious-to-set-up processing pipelines. TractSeg is a novel convolutional neural network-based approach that directly segments tracts in the field of fiber orientation distribution function (fODF) peaks without requiring tractography, image registration or parcellation. We demonstrate in 105 subjects from the Human Connectome Project that the proposed approach is much faster than existing methods while providing unprecedented accuracy. The code and data are openly available at https://github.com/MIC-DKFZ/TractSeg/ and https://doi.org/10.5281/zenodo.1088277, respectively.
1. Introduction
White matter tract segmentation supports tractometry but existing virtual-dissection, ROI-based, and clustering-based workflows are complex, time-consuming, computationally expensive, and sensitive to registration or user-dependent steps. TractSeg addresses this gap through direct FCNN-based segmentation of fODF peaks without registration, parcellation, tractography, or clustering.
- Accurate tract segmentation is required for tractometry that characterizes healthy brains and identifies abnormal morphology in diseased brains.
- Virtual dissection manually extracts anatomically defined streamlines using inclusion and exclusion ROIs, making the process time-consuming and subject to limited reproducibility.
- ROI- and clustering-based methods require pipelines involving registration, tractography, parcellation, or clustering, with substantial setup, computation, tuning, and possible error propagation.
- Direct segmentation avoids intermediate streamline generation by producing complete tract segmentations directly from input images.
- TractSeg directly segments white matter tracts from fODF peaks using an FCNN and is designed to be accurate, fast, easy to set up, and independent of registration, parcellation, tractography, and clustering.
2. Materials and Methods
TractSeg uses orientation-aware encoder-decoder FCNNs to transform compact fODF-peak representations into multi-tract probability maps. The method was trained and evaluated on 72 reference tracts from 105 HCP subjects using high-quality and reduced-quality diffusion MRI data.
- Segmentation pipeline: The pipeline extracts three dominant diffusion directions per voxel, predicts tract probabilities from axial, coronal, and sagittal views, and averages orientation-specific outputs into final segmentations.A first FCNN produces orientation-specific tract maps; a second FCNN refines their combinations before mean merging.
- Input representation: The method uses a condensed three-direction-per-voxel representation to reduce acquisition dependence and avoid the high channel count and input/output demands of raw high-angular-resolution data.
- Convolutional Neural Network: The network maps 144x144 2D inputs with 9 fODF-peak channels to 72 tract-probability channels and thresholds probabilities at 0.5 for binary segmentations.Sigmoid outputs support multilabel segmentation, allowing several tracts to share a voxel.
- Convolutional Neural Network: 2D slices improve memory efficiency while sampling axial, coronal, and sagittal orientations preserves information from the third dimension.
- Training: Training sampled three MRtrix peak types, multiple orientations, and randomized augmentations across 500 epochs, totaling 4,536,000 slices.
- Reference segmentations: Reference segmentations covered 72 major white matter tracts and were generated through a semi-automatic dissection pipeline involving tractography and refinement.
3. Experiments and results
Across quantitative and qualitative evaluations, TractSeg achieved strong tract segmentation accuracy, remained comparatively robust on lower-quality clinical data, and substantially reduced runtime. Performance varied by tract difficulty and fusion strategy, with thin or complex tracts remaining challenging.
- Quantitative evaluation: 9 Dice points: TractSeg significantly outperformed the second-best method on HCP Quality; the margin was 14 Dice points on Clinical Quality.Both comparisons used mean Dice over 72 reference tracts and were significant at p<0.001.
- Quantitative evaluation: All tracts exceeded 0.75 Dice except the fornix and anterior commissure, whose thin shapes reduced performance.Incomplete reference tract reconstructions also produced lower Dice scores despite sometimes better model segmentations.
- Qualitative evaluation: On representative HCP tracts, TractSeg produced adequate reconstructions, while competing methods showed tract-specific omissions, oversegmentation, or false positives.The examples covered the corticospinal tract, anterior commissure, and inferior occipito-frontal fascicle.
- Qualitative evaluation: Clinical-quality segmentations were lower quality across methods, but TractSeg appeared less affected than most benchmarks and registration methods often included gray matter or outside-brain regions.Streamline-based methods showed increased false positives on the Clinical Quality dataset.
- Fusion strategies: Mean fusion slightly outperformed majority voting on average, while the second FCNN outperformed mean fusion on Clinical Quality by 1.9 Dice points.On HCP Quality, mean fusion was 0.2 Dice points better than the second FCNN; across Clinical Quality tracts, the second FCNN was especially stronger on the fornix.
- Runtime: 81x faster on HCP Quality and 535x faster on Clinical Quality: TractSeg substantially reduced runtime relative to reference methods.The segmentation itself took about 1 minute, or 0.5 minutes with mean fusion; tractography-dependent methods exceeded 950 minutes.
4. Discussion and Conclusion
TractSeg provides accurate direct white matter tract segmentation while avoiding several complex pipeline stages and remaining robust to reduced image quality. Its evaluation and open resources support practical use, but conclusions depend on imperfect reference data and supervised training.
- Overview: TractSeg achieved unprecedented accuracy across 72 tracts in 105 HCP subjects and was less affected by reduced-resolution clinical-quality data.The evaluation covered both high-quality and downsampled, more clinical-like datasets.
- Limitations: The reference tracts are not real ground truth, with possible false positives, incomplete reconstructions, and variation in detailed anatomical definitions.These limitations affect interpretation of accuracy comparisons, although the authors describe the dataset as a strong in-vivo approximation of white matter anatomy.
- Method and robustness: Robustness to differing image resolutions was attributed to heavy augmentation, multi-view ensembling, and CNN tolerance of noisy data.The three-view predictions can be fused by Mean, Majority Voting, or an optional Second FCNN.
- Limitations: TractSeg depends on the availability and quality of supervised training data, and scanner or acquisition differences can reduce performance through domain shift.Heavy augmentation can reduce, but does not eliminate, this domain-shift concern.
- Overview: TractSeg avoids atlas registration, tractography, parcellation, and clustering, reducing the processing steps required by existing pipelines.Existing pipelines also require additional runtime, installation, setup, and quality-control steps.
- Resources: The 72-tract HCP reference dataset and pretrained TractSeg package are openly available to support standardized method evaluation and reuse.The authors provide the dataset and software through Zenodo and GitHub.