Supervised Learning of Tractograms Alignment Driven by Fibers Correspondence
Poster No:
1490
Submission Type:
Abstract Submission
Authors:
Gabriele Amorosino1, Mattias Heinrich2, Paolo Avesani3
Institutions:
1The University of Texas at Austin, Austin, TX, 2University of Luebeck, Luebeck, Luebeck, 3Fondazione Bruno Kessler, Trento, Trento
First Author:
Co-Author(s):
Introduction:
Accurate alignment of tractograms is crucial for comparing WM structures across individuals and supporting clinical studies in neuroscience. Traditional tractogram alignment methods rely on volumetric registration, such as using T1-weighted MRI images, which lack sensitivity to the orientation and continuity of WM fibers. These approaches often fail to capture the complex fiber structures, limiting their ability to perform accurate alignments. In this study, we introduce DGTA (Deep learning-based Geometric Tractogram Alignment), the first deep learning method designed to directly align tractograms using tractography fibers.
Methods:
To address the task of tractogram alignment, we developed DGTA, a geometric deep learning model that employs graph convolutional networks (GCNs) and loopy belief propagation (LBP) [1,2]. The model directly aligns tractograms by learning displacement fields between corresponding fiber bundles across subjects. Tractograms were represented as point clouds, and graph-based encoding strategies were applied to capture the relationships between fiber points. Specifically, we tested three encoding strategies: i) pts (k-nearest neighbors), ii) poly (fiber polyline structure), and iii) trk (neighboring fibers), with the poly strategy showing the most effective results.
For training and evaluation, we used the TractoInferno dataset [3], which contains whole-brain tractograms and 29 manually curated fiber bundles from 20 subjects. The dataset was split into 16 training subjects (240 pairwise comparisons) and 4 test subjects (12 pairwise comparisons). We employed a novel template-based subsampling strategy to reduce the computational complexity by selecting 1000 fibers (16,000 points) per tractogram. Ground truth was established using fiber bundle skeletons, which represent the average trajectories of fibers within each bundle, allowing us to compute pseudo-ground truth displacement fields for supervised learning.
The model was trained over 70 epochs with an initial learning rate of 0.01, later reduced to 0.001. Comparative experiments were conducted between DGTA and state-of-the-art volumetric registration methods, including ANTs SyN, with alignment performance assessed using metrics such as mean square error (MSE), Hausdorff distance (HD), and Linear Assignment Problem Distance (LAPD), which emphasizes fiber density alignment.
For training and evaluation, we used the TractoInferno dataset [3], which contains whole-brain tractograms and 29 manually curated fiber bundles from 20 subjects. The dataset was split into 16 training subjects (240 pairwise comparisons) and 4 test subjects (12 pairwise comparisons). We employed a novel template-based subsampling strategy to reduce the computational complexity by selecting 1000 fibers (16,000 points) per tractogram. Ground truth was established using fiber bundle skeletons, which represent the average trajectories of fibers within each bundle, allowing us to compute pseudo-ground truth displacement fields for supervised learning.
The model was trained over 70 epochs with an initial learning rate of 0.01, later reduced to 0.001. Comparative experiments were conducted between DGTA and state-of-the-art volumetric registration methods, including ANTs SyN, with alignment performance assessed using metrics such as mean square error (MSE), Hausdorff distance (HD), and Linear Assignment Problem Distance (LAPD), which emphasizes fiber density alignment.
·DGTA architecture.
Results:
Our empirical analyses showed that DGTA significantly outperformed traditional volumetric registration methods, including ANTs SyN [4], in terms of tractogram alignment accuracy. DGTA achieved superior results across all alignment metrics, particularly in aligning dense fiber regions. When evaluated using the HD and RMSE metrics, DGTA consistently produced lower alignment errors across all test subjects compared to volumetric methods. The polyline-based graph encoding strategy (poly) was the most effective, outperforming both pts and trk strategies in preserving the intrinsic fiber structure during alignment.
In terms of whole-brain tractogram alignment, DGTA reduced the Bundle Minimum Distance (BMD) metric [5] compared to the baseline ACPC alignment, demonstrating that the method effectively captures fiber geometry at both global and local scales. However, BMD's correlation with Dice Score indicated that it is more reflective of volumetric alignment rather than fiber density. The LAPD metric while captures the alignment of dense fiber regions, confirmed that DGTA provides better alignment of these critical areas compared to ANTs SyN.
In terms of whole-brain tractogram alignment, DGTA reduced the Bundle Minimum Distance (BMD) metric [5] compared to the baseline ACPC alignment, demonstrating that the method effectively captures fiber geometry at both global and local scales. However, BMD's correlation with Dice Score indicated that it is more reflective of volumetric alignment rather than fiber density. The LAPD metric while captures the alignment of dense fiber regions, confirmed that DGTA provides better alignment of these critical areas compared to ANTs SyN.
·Empirical results.
Conclusions:
DGTA represents a significant advancement in tractography alignment by utilizing fiber geometry to achieve more accurate results compared to voxel-based methods. The model's scalability, combined with its improved performance in aligning dense fiber regions, makes it a valuable tool for neuroimaging research and clinical applications, particularly in studies of brain connectivity and neurodegenerative diseases.
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 2
Image Registration and Computational Anatomy 1
Keywords:
Machine Learning
Tractography
White Matter
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Hansen, L., & Heinrich, M. (2021a). Geometric Deep Learning model for point cloud alignment.
3. Poulin, P., et al. (2022). TractoInferno dataset.
4. Avants, B. B., et al. (2011). ANTs SyN algorithm for non-linear registration of brain volumes.
5. Garyfallidis, E., et al. (2015). Bundle Minimum Distance metric.