Abstract
Diffusion magnetic resonance imaging (dMRI) fiber tractography is a powerful tool for investigating human white matter connectivity in vivo. However, it is prone to false positive and false negative results, making interpretation of the tractography result difficult. Optimal tractography must begin with an accurate description of the subvoxel white matter fiber structure, include quantification of the uncertainty in the fiber directions obtained, and quantify the confidence in each reconstructed fiber tract. This paper presents a novel and comprehensive pipeline for fiber tractography that meets these requirements. The subvoxel fiber geometry is described in detail using a technique that allows not only for straight crossing fibers but for fibers that curve and splay. This technique is repeatedly performed within a residual bootstrap statistical process in order to efficiently quantify the uncertainty in the subvoxel geometries obtained. A robust connectivity index is defined to quantify the confidence in the reconstructed connections. The tractography pipeline is demonstrated in the human brain.
Introduction
Diffusion MRI tractography is the primary method for mapping the structural connectivity of the living human brain. Despite its utility, the field has long struggled with accuracy. Early methods like Diffusion Tensor Imaging (DTI) were limited by their inability to resolve crossing fibers within a single voxel. While High Angular Resolution Diffusion Imaging (HARDI) addressed simple crossings, it remained challenged by "ambiguous" geometries such as fanning (splaying) or curving fibers.
In standard orientation distribution functions (ODFs), a voxel containing fibers that fan out can appear identical to one where fibers simply curve. If a tracking algorithm cannot distinguish these structures, it may terminate prematurely or follow biologically incorrect paths. Furthermore, most tractography lacks a reliable metric to quantify the confidence of a specific reconstructed connection.
Methodology
The proposed pipeline addresses these limitations through a multi-stage approach that incorporates complex geometry modeling and statistical uncertainty estimation.
1. Residual Bootstrap Framework
To estimate noise without the need for multiple expensive MRI scans, the researchers utilize a "residual bootstrap" method. By fitting a Spherical Harmonics model to a single scan, they calculate residuals (errors) and randomly reshuffle them to simulate 100 independent versions of the dataset. This allows for the quantification of uncertainty in fiber orientation.
2. Curve Inference and Geometry Modeling
For each bootstrap iteration, a 3D curve inference algorithm is applied. This algorithm uses local differential geometry—specifically "co-helicity"—to analyze a voxel's neighborhood. It determines whether the underlying structure represents a single fiber, a crossing, or a fanning system.
3. Polarity-Aware Tracking and Connectivity
When a voxel is identified as "fanning," the algorithm determines its polarity (the direction of spread), allowing streamlines to propagate accurately through the fan. To quantify confidence, the pipeline uses a "weakest link" approach: the connectivity index of an entire tract is determined by the confidence of its most uncertain segment, rather than a simple streamline count.
Key Insights: Fanning vs. Curving
A central finding of this work is that fanning cannot be identified by looking at a single voxel in isolation. By examining the neighborhood, the algorithm can distinguish between a splaying structure (fan) and a single bending path (curvature). This distinction is vital for maintaining tracking accuracy in regions where white matter tracts naturally disperse as they approach the cortex.
Results and Validation
The pipeline was validated using in vivo human data across various protocols, including high-resolution (99 directions) and standard clinical-grade (64 directions) scans. Key results include:
- Bootstrap Reliability: Scan-rescan validation showed that the 95% confidence intervals generated from a single scan correctly predicted fiber directions in subsequent scans with high accuracy in major white matter tracts.
- Improved Cortical Mapping: Modeling fanning prevented the typical "drop-off" in confidence seen as fibers approach the brain's cortex, where they naturally splay.
- Anatomical Accuracy: The system successfully mapped complex regions like Broca’s area and the internal capsule, identifying clear fanning patterns that simpler algorithms typically miss.
Conclusion
By moving beyond simple crossing-fiber models, this pipeline provides a more biologically plausible map of brain connectivity. While the computational cost is currently high—requiring significant processing time for 100 bootstrap repetitions—the method is highly parallelizable. Future work will focus on optimizing these algorithms and extending the model to handle even more complex scenarios, such as intersecting fanning systems.
Summary updated