PaperlayerView exact revision on arXiv

Source-linked AI summary

Combined tract segmentation and orientation mapping for bundle-specific tractography

Jakob Wasserthal, Peter Neher, Dusan Hirjak, Klaus H. Maier-Hein

arXiv:1901.10271v2cs.CV

TL;DR

Manual dissection of white-matter tracts is difficult to scale and reproduce, motivating automated bundle-specific tractography. The paper combines tract orientation maps with tract and endpoint segmentation and probabilistic tracking. Across 72 bundles and varied datasets, it reports higher accuracy and faster runtime than seven benchmark methods, while generalizing to unseen acquisitions and pathologies.

  • Problem

    Manual dissection of white-matter tracts is time-consuming, requires expertise, and is difficult to reproduce; automated methods often involve cumbersome processing.

  • Method

    The method combines learned tract orientation maps with tract-outline and start/end-region segmentation and probabilistic tracking.

  • Results

    The approach significantly outperformed reference methods by an average of 14 Dice points on HCP Quality and by a large margin on lower-quality datasets.

  • Takeaways & Limitations

    The approach produces accurate bundle-specific tractograms without whole-brain tractography, registration, parcellation, clustering, or manual dissection.

  • Takeaways & Limitations

    Performance depends on the availability and quality of supervised training data, and acquisition differences introduce domain shift that can reduce performance.

Abstract

from arXiv · show

While the major white matter tracts are of great interest to numerous studies in neuroscience and medicine, their manual dissection in larger cohorts from diffusion MRI tractograms is time-consuming, requires expert knowledge and is hard to reproduce. In previous work we presented tract orientation mapping (TOM) as a novel concept for bundle-specific tractography. It is based on a learned mapping from the original fiber orientation distribution function (FOD) peaks to tract specific peaks, called tract orientation maps. Each tract orientation map represents the voxel-wise principal orientation of one tract. Here, we present an extension of this approach that combines TOM with accurate segmentations of the tract outline and its start and end region. We also introduce a custom probabilistic tracking algorithm that samples from a Gaussian distribution with fixed standard deviation centered on each peak thus enabling more complete trackings on the tract orientation maps than deterministic tracking. These extensions enable the automatic creation of bundle-specific tractograms with previously unseen accuracy. We show for 72 different bundles on high quality, low quality and phantom data that our approach runs faster and produces more accurate bundle-specific tractograms than 7 state of the art benchmark methods while avoiding cumbersome processing steps like whole brain tractography, non-linear registration, clustering or manual dissection. Moreover, we show on 17 datasets that our approach generalizes well to datasets acquired with different scanners and settings as well as with pathologies. The code of our method is openly available at https://github.com/MIC-DKFZ/TractSeg.

1. Introduction

The paper addresses false positives and reproducibility challenges in bundle-specific tractography by extending tract orientation mapping with segmentation and probabilistic tracking. It evaluates the resulting approach across diverse bundles and datasets.

  • Manual tract dissection is time-consuming, expert-dependent, and difficult to reproduce, while overlapping tracts make tractography susceptible to false positives.
  • Existing automated approaches commonly rely on whole-brain tractography, registration, clustering, or other tedious processing steps.
  • Tract orientation mapping learns tract-specific peak images in which each voxel represents the local orientation of one tract.
  • The proposed extension combines tract orientation maps with tract-outline and start/end-region segmentation and a more sensitive probabilistic tracking algorithm.
  • The approach is reported as easy to set up and fast to run without affine or elastic registration, parcellation, or clustering.

2. Materials and Methods

The method uses FOD peak directions as a compact input representation for three related neural-network tasks. This design supports variation in MRI acquisition while avoiding the dimensionality and memory burden of raw diffusion inputs.

  • A common U-Net architecture receives FOD peaks and learns tract segmentation, start/end-region segmentation, or tract orientation mapping through different training targets.
  • The network input comprises the three principal fiber directions per voxel, requiring nine channels rather than one channel per gradient orientation and b-value.
  • FOD peaks are estimated from diffusion data using multi-shell multi-tissue constrained spherical deconvolution and peak extraction in MRtrix.

2.2. Model

The model uses a 2D U-Net-based architecture to predict tract masks, start/end regions, and tract orientation maps, which are then combined through probabilistic tracking and filtering.

  • A 2D encoder-decoder architecture inspired by U-Net uses deep supervision to improve gradient flow, reduce training time, and slightly improve results.
  • The network processes 144 × 144 images with nine input channels and outputs tract-specific channels for masks, start/end regions, or orientation vectors.
  • CSD-derived FOD directions feed three U-Nets that predict a tract orientation map, tract mask, and start/end-region mask for each tract.
  • Streamlines are retained only when they remain within the tract mask and end in the predicted start/end regions, producing one tractogram per tract.
  • For tract segmentation and start/end segmentation, predictions from three slice orientations are merged, whereas tract orientation mapping uses a single orientation because three-view inference slightly worsened results.

2.3. Training

Training uses augmentation and normalization to improve robustness across acquisition conditions. Segmentation and orientation models use task-specific losses and standard supervised optimization.

  • Segmentation models use sigmoid outputs and binary cross-entropy loss for target masks across classes.
  • The tract orientation mapping model uses linear outputs and cosine similarity loss to learn tract peak directions.
  • Adamax optimization uses a learning rate of 0.001, batch size 47, and learning-rate reduction after 20 validation epochs without improvement.
  • Heavy augmentation applies random rotations and elastic deformations sampled from uniform distributions to improve generalizability.
  • Training samples are normalized to zero mean and unit variance, and multiple peak-image variants are used to support different MRtrix CSD inputs.
  • Models are trained on 144 × 144 images matching the 1.25 mm HCP resolution, with downsampling and upsampling used to simulate blurred lower-resolution inputs.

2.4. Data

The study uses 72 reference tract delineations from 105 HCP subjects, preprocessing them into masks for tract, start/end regions, and orientation mapping. Clinical-quality and phantom datasets extend evaluation to lower-quality and simulated acquisition conditions.

  • 72 major white matter tracts from 105 HCP subjects provide the reference delineations for training and evaluation.
  • Reference streamlines are converted into binary tract masks and a unified start/end-region mask for the three learning tasks.
  • DBSCAN separates the combined start/end-region mask into two clusters, avoiding manual separation when regions are close.
  • Clinical Quality data are downsampled to 2.5 mm isotropic resolution with 32 b = 1000s/mm2 volumes while retaining HCP reference tracts.
  • Reference delineations from another scanner were unavailable, while low-quality expert tract delineations were considered too ambiguous for detailed evaluation.
  • The Phantom dataset simulates 21 held-out subjects at 2.5 mm resolution with 32 directions and randomly selected acquisition artefacts.

2.5. Bundle-specific tractography

The method compares three ways to use tract orientation maps and selects direct tracking on TOMs. A constrained probabilistic tracker improves sensitivity while masks limit spurious streamlines, especially in low-quality data.

  • Three TOM tracking strategies use direct TOM tracking, nearest original peaks, or a weighted mean between TOM and original orientations.
  • Direct tracking on TOMs was selected because it provided the best balance of high sensitivity and high specificity, particularly on low-quality data.
  • Tracking on original signal peaks was insufficient: deterministic tracking lacked sensitivity, whereas probabilistic tracking produced many false positives.
  • A Gaussian distribution centered on each TOM orientation enables probabilistic sampling that can reconstruct small branchings missed by deterministic tracking.
  • The tract mask and start/end-region mask constrain probabilistic tracking, while a fixed standard deviation of 0.15 keeps false positives relatively small.
  • Higher sampling variation, such as a standard deviation of 0.3, produces many spurious fibers, whereas 0.15 remains consistent with deterministic tracking.

2.6. Reference methods

The evaluation compares the proposed approach with automatic delineation, atlas, clustering, registration, and orientation-based reference methods. The comparisons include both performance measures and method-specific applicability constraints.

  • The study compares automatic delineation methods using DICE for segmentation and voxel-wise angular error for orientation quality.
  • The benchmark includes TractQuerier, RecoBundles, streamline atlas, custom atlas registration, FSL atlas registration, and multiple-mask registration.
  • Peak atlas and best original peak provide additional comparisons for voxel-wise tract orientation quality.
  • The compared approaches include ROI-based and clustering-based methods, often relying on whole-brain tracking, streamline processing, or registration.
  • The FOD-registration atlas method was not applicable to Phantom data because simulated and template FODs were insufficiently similar for meaningful registration.
  • The best-original-peak comparison uses reference peaks and therefore is not a fair direct comparison, but estimates the quality of the original peaks.

3. Experiments and results

Across datasets and evaluation settings, the proposed method achieved strong segmentation and orientation performance, produced spatially coherent reconstructions, generalized to pathological cases, and ran substantially faster than reference methods.

  • Segmentation performance: 14, 18, and 22 Dice points were the proposed method’s average advantages over reference methods on HCP Quality, Clinical Quality, and Phantom datasets, respectively.The improvements were significant at p < 0.01, and the method was less affected by quality loss.
  • Orientation performance: The proposed method’s angular error increased by only 1 degree from HCP Quality to Clinical Quality data.RecoBundles and TractQuerier had slightly better orientation errors on HCP Quality but worsened significantly on lower-quality datasets.
  • Qualitative evaluation: On the HCP Quality test set, the proposed method yielded accurate and spatially coherent reconstructions for CST, CA, and IFO.The comparison used a representative subject selected by mean-dataset Dice performance.
  • Qualitative evaluation: Reference methods showed characteristic failure modes including oversegmentation, premature termination, endpoint-driven wrong turns, and registration-induced tract shifts.These issues affected RecoBundles, TractQuerier, and SLAtlas in the qualitative comparisons.
  • Generalization to pathologies: The proposed method reconstructed affected tracts in Alzheimer’s, multiple sclerosis, and reduced-brain-volume cases despite distortions, lesions, or pathology-related signal loss.It reconstructed CST and OR with enlarged ventricles, fibers through MS lesions, and T POSTC fibers in the postcentral gyrus despite reduced brain volume.

4. Discussion and conclusion

The proposed method achieves strong accuracy and runtime across 72 tracts and generalizes to diverse datasets, while remaining subject to reference-data, preprocessing, and supervised-learning limitations.

  • 72 tracts were evaluated on HCP, reduced-quality, phantom, and unseen datasets, with reported accuracy and runtime advantages over seven benchmark methods.The method was also evaluated on datasets with different scanners, settings, and pathologies.
  • The training and evaluation tract delineations are diffusion-based approximations rather than real ground truth, although phantoms provide available ground truth.The HCP dataset was manually inspected and described as one of the best existing in vivo approximations at this cohort size.
  • Phantom Dice scores were lower than Clinical Quality scores, but the method still produced complete reconstructions with only minor errors.The lower scores were attributed to simulated artifacts and a larger domain shift from training data.
  • Reference-method comparisons have limitations, including potential TractQuerier bias, only five RecoBundles reference subjects, and unoptimized default settings.The authors state that these limitations alone cannot explain the large accuracy gap between their method and the reference methods.
  • Preprocessing uses multiple non-trivial parameter choices whose influence on final results was not systematically evaluated.Changing these parameters could significantly alter the final results.
  • Because the method uses supervised learning, its performance depends on training-data availability and quality, with domain shift reducing performance.Heavy data augmentation reduced this shift, and experiments nevertheless produced anatomically plausible results on most unseen datasets and pathologies.

List of all tracts

The model supports 72 named white-matter tracts, including association, projection, commissural, limbic, and cerebellar pathways.

  • The supported tract list includes AF, ATR, CA, CC subdivisions, CG, CST, MLF, FPT, FX, ICP, IFO, and additional listed tracts.The passage begins the complete 72-tract inventory and includes corpus callosum subdivisions CC 1 through CC 7.

RecoBundles default parameters

The supplied material documents reference-method registration and recognition parameters alongside figures describing the proposed network, evaluations, runtime, crossings, and sampling behavior.

  • RecoBundles default parameters: Whole-brain streamline-based registration uses clustering threshold 5, local linear registration, similarity and affine transforms, progressive optimization, and a small matrix.The listed settings also specify no registration metric.
  • RecoBundles default parameters: Registration optimization allows 150 iterations and initializes QuickBundles with 50,000 random streamlines.
  • RecoBundles default parameters: Recognize-bundles settings use clustering threshold 15, reduction threshold 10, and MDF as the reduction distance.
  • Proposed U-Net architecture: The proposed U-Net adds convolutional layers in the upsampling path to improve gradient flow through deep supervision.Blue boxes denote multi-channel feature maps, white boxes copied feature maps, and gray labels encode channels and spatial size.
  • Evaluation figures: Figures report qualitative results across healthy and pathological datasets, phantom comparisons, runtime for segmentation and orientation estimation, and resolution of a three-tract crossing.The figures also compare model dimensionality and inputs, endpoint filtering, Dice scores across 72 tracts, and sampled-orientation dispersion.
  • Orientation sampling: The orientation histogram compares original FOD sampling with fixed-Gaussian sampling using standard deviation 0.15 and reports lower dispersion for the fixed distribution.
Loading 1901.10271v2…