Deep Learning the Shape of the Brain Connectome

Haocheng Dai, Martin Bauer, P. Thomas Fletcher, Sarang C. Joshi

Abstract

Mapping the structural connectivity of the human brain, or tractography, remains a fundamental challenge in neuroimaging. This paper introduces a novel deep learning framework that represents white matter as a Riemannian manifold. By employing Convolutional Encoder-Decoder Neural Networks (CEDNNs) to estimate the underlying geometric metric, the proposed method accurately resolves complex fiber crossings within a geodesic framework, significantly outperforming traditional streamline and geodesic approaches.

Introduction

Diffusion-weighted magnetic resonance imaging (DWI) is the primary tool for inferring neural tract paths by measuring water molecule diffusion. However, traditional tractography faces significant hurdles: noise, high curvature of pathways, and the "crossing fiber" problem where multiple pathways intersect within a single voxel. Geodesic tractography offers a sophisticated alternative by treating the brain as a curved space where tracts are geodesics. The core challenge lies in the inverse problem: calculating a Riemannian metric that aligns local geometry with observed fiber directions.

Methods

The authors frame metric estimation as solving a system of highly nonlinear Partial Differential Equations (PDEs). They utilize a Convolutional Encoder-Decoder Neural Network (CEDNN) architecture, incorporating "Dense Blocks" to enhance information flow and spatial detail recovery.

The CEDNN architecture utilizes a physics-informed loss function and Rodrigues’ rotation formula to efficiently estimate an SPD Riemannian metric.

The network is trained in a self-supervised manner using a physics-informed loss function that minimizes the residual of the geodesic PDE. To ensure the estimated metric is a Symmetric Positive-Definite (SPD) matrix, the authors implement an efficient eigencomposition approach using Rodrigues’ rotation formula, which significantly accelerates training compared to standard matrix exponential methods.

Results

The framework was validated using both synthetic datasets (e.g., a "braid" pattern for crossing fibers) and real-world data from 38 subjects in the Human Connectome Project (HCP).

Qualitative and quantitative comparison showing the proposed method's superior alignment with white matter pathways and lower error rates compared to baseline metrics.

Key findings include:

  • Superior Alignment: The proposed method significantly outperformed state-of-the-art baselines (inverted, adjugate, and conformal metrics) in aligning geodesics with actual white matter pathways.
  • Resolving Fiber Crossings: In synthetic tests, the CEDNN successfully recovered crossing pathways without the confounding or merging typical of previous geodesic methods.
  • Efficiency: A full 3D brain metric can be estimated in under 30 minutes on a single consumer-grade GPU.
  • Robustness: CEDNNs proved more effective than standard Physics-Informed Neural Networks (PINNs) for structured grid data and showed high resilience at tissue boundaries.

Discussion and Conclusion

While the current model requires per-subject training due to limited generalization to unseen data, it represents a significant leap in geodesic tractography. By enabling the optimization of a single Riemannian metric compatible with multiple vector fields, this approach successfully addresses the crossing fiber problem. This robust metric estimation paves the way for "Geometrical Statistics," allowing for the construction of statistical brain atlases and precise quantification of white matter changes in neurodegenerative diseases.

Reference: Dai, H., Bauer, M., Fletcher, P. T., & Joshi, S. (2023). Modeling the Shape of the Brain Connectome via Deep Neural Networks. arXiv:2203.06122v2.

Summary updated

BibTeX
@inproceedings{dai2023deep,
  author = {Haocheng Dai and Martin Bauer and P. Thomas Fletcher and Sarang C. Joshi},
  title = {{Deep Learning the Shape of the Brain Connectome}},
  booktitle = {IPMI},
  year = {2023},
  doi = {10.1007/978-3-031-34048-2\_23},
}
Figure