Introduction
Diffusion MRI tractography is a cornerstone of modern neuroimaging, allowing researchers to non-invasively map the complex white matter architecture of the human brain. However, the process is notoriously prone to "false positives"—spurious fibers that are isolated, poorly aligned with known anatomical structures, or artifacts of the reconstruction algorithm. Manually cleaning these datasets is labor-intensive and subjective. This paper, presented at the 2016 Organization for Human Brain Mapping (OHBM) meeting, introduces an automated, mathematically grounded method for "cleaning" tractography data using Fiber to Bundle Coherence (FBC).
Methodology: Fiber to Bundle Coherence (FBC)
The core of the proposed approach is the FBC measure, which leverages advanced differential geometry to identify and prune incoherent fibers. The method operates through a series of sophisticated steps:
- 5D Lifting: Instead of treating fibers as simple 3D curves, they are "lifted" into a 5D space that accounts for both 3D position and 2D local orientation (tangent).
- Lie Group Geometry: This 5D space is modeled using the rigid body motion group $SE(3)$. A sub-Riemannian metric is applied to penalize "sideways" movement, reflecting the biological reality that nerve fibers bend and twist but do not move laterally.
- Kernel Density Estimation (KDE): The density of fibers is calculated in this 5D space. A fiber surrounded by many others with similar orientations receives a high density score.
- LFBC and RFBC:
- Local FBC (LFBC): A point-wise measure of how well a fiber segment aligns with its neighbors.
- Relative FBC (RFBC): A global measure for the entire fiber, calculated by taking the minimum moving average of the LFBC along the fiber's length.
Implementation and Performance
To make this complex mathematical theory accessible, the authors implemented the tool within the open-source DIPY (Diffusion Imaging in Python) library. The implementation utilizes Cython and OpenMP for multithreading, KD-trees for fast orientation matching, and pre-computed lookup tables for geometric kernels. These optimizations provide a considerable speedup, making kernel density estimation feasible for large-scale datasets.
Results: Pruning the Optic Radiation
The method was evaluated using the Stanford HARDI dataset (160 orientations, b-value = 3000 s/mm²). Constrained Spherical Deconvolution (CSD) was used to reconstruct the optic radiation, which connects the visual cortex (V1) to the lateral geniculate nucleus.
Initial tractography results were "fuzzy," containing many incoherent fibers. By applying a threshold (removing fibers with an RFBC lower than 0.3), the authors successfully pruned the dataset to produce a much cleaner and more anatomically plausible bundle.
Discussion and Future Work
The fundamental insight of FBC is that biological white matter travels in organized bundles; therefore, "lonely" or misaligned fibers are statistically likely to be errors. While the results are qualitatively superior, several areas for future work remain:
- Threshold Selection: Determining an "optimal" threshold across different brain regions or subjects remains an open question.
- Quantitative Validation: Further studies are needed to validate the method against "ground truth" synthetic phantoms.
- Integration: The FBC module is designed for easy integration into standard processing pipelines to improve the reliability of connectivity studies.
By moving sub-Riemannian geometry from theory into a practical, open-source tool, the authors provide a robust mechanism for improving the anatomical accuracy of reconstructed brain pathways.
Summary updated