Anatomy-to-tract mapping infers white matter pathways without diffusion streamline propagation

Yee Fan Tan, Khoi Minh Huynh, Siyuan Liu, Raphaël C.‐w. Phan, Chee‐Ming Ting, Pew‐Thian Yap

TL;DR

This paper introduces Anatomy-to-Tract Mapping (ATM), a novel deep learning framework that reconstructs white matter pathways directly from anatomical T1-weighted MRI scans, bypassing the need for diffusion MRI data. ATM leverages the high quality of anatomical scans to generate anatomically plausible, subject-specific streamlines, overcoming limitations of traditional diffusion-based methods.

What Problem Does This Paper Address?

Mapping white matter pathways in the brain is crucial for understanding brain structure and function, and for diagnosing neurological disorders. The standard technique for this is diffusion MRI (dMRI) tractography, which infers fiber directions from how water molecules diffuse in brain tissue. However, dMRI tractography faces significant challenges:

  • Signal-to-Noise Ratio (SNR) and Resolution: dMRI data often has limited SNR and spatial resolution, especially in clinical settings, making it difficult to accurately estimate fiber orientations.
  • Complex Fiber Configurations: When multiple fiber bundles cross, kiss, or bend within a single voxel, it becomes challenging to accurately determine their orientations, leading to errors in reconstructed pathways.
  • Bottleneck Problem: In narrow regions where many fiber populations converge, orientation ambiguities can lead to false positive streamlines.

These limitations hinder the reliability and accuracy of dMRI tractography, motivating the search for alternative approaches.

Methods

The core innovation of this paper is the Anatomy-to-Tract Mapping (ATM) framework. Unlike traditional methods that rely on diffusion signals, ATM uses anatomical T1-weighted (T1w) MRI scans as its primary input. Here's how it works:

  • Anatomy-Driven Learning: ATM leverages a deep learning model trained on paired T1w images and corresponding ground truth tractograms. It learns to associate anatomical features (like cortical folding and tissue boundaries) with white matter pathways.
  • No Streamline Propagation: Instead of propagating streamlines point-by-point based on local diffusion information, ATM directly generates complete, subject-specific bundles. This avoids the inherent limitations of streamline propagation in complex regions.
  • Generative Approach: ATM uses a variational autoencoder (VAE) architecture. It encodes anatomical features from the T1w image and then decodes them into anatomically plausible streamlines. This generative process allows it to synthesize realistic pathways.
  • Filtering and Trimming: After initial generation, streamlines are refined using T1w-derived tissue masks and white matter surfaces to remove implausible or extraneous pathways.
  • Population-Level Inference: ATM can also generate population-averaged bundles by processing averaged anatomical scans, providing a generalized view of white matter pathways.

Datasets

  • TractoInferno Dataset: This large-scale, open-source dataset was used for both training and evaluating ATM. It includes T1w images and ground truth streamlines for 30 white matter bundles across 284 subjects. The ground truth streamlines were generated using multiple established tractography algorithms and manual curation.
  • External Clinical Dataset: ATM was also evaluated on T1w scans from an external clinical dataset of low-field MRI, demonstrating its applicability to data with potentially lower quality.

Key Insights and Ideas

  • Anatomy as a Sufficient Cue: The most striking insight is that high-quality anatomical MRI alone can provide enough information to infer complex white matter pathways, challenging the long-held reliance on diffusion MRI for this task.
  • Bypassing dMRI Limitations: By avoiding diffusion signal processing, ATM sidesteps issues related to SNR, resolution, and the "bottleneck problem" that plague traditional tractography.
  • Subject-Specific Reconstruction: ATM's ability to generate anatomically plausible, subject-specific bundles directly from individual T1w scans is a significant advancement over population-based atlases or methods that struggle with individual variability.
  • Bridging Spatial Scales: The work suggests that tractography can be viewed as a multimodal phenomenon, bridging the microstructural information from dMRI with the macrostructural information from T1w MRI.

Key Results

ATM demonstrated strong performance across various metrics when compared to established diffusion-based methods (MRtrix with BundleSeg, SCIL atlas warping) and a population-based ATM approach:

  • Bundle Similarity: ATM achieved high bundle adjacency (BA) scores, indicating excellent shape similarity with ground truth.
  • Coverage: Competitive Dice, overlap, and overreach scores were observed, suggesting good spatial coverage and accuracy.
  • Validity: ATM generated a high proportion of valid streamlines, indicating that the reconstructed pathways adhere to expected anatomical trajectories.
  • Orientation Alignment: High angular correlation coefficients (ACCs) between predicted and ground truth tract orientation density maps were achieved, demonstrating accurate fiber direction inference.
  • Geometric Fidelity: ATM showed lower absolute percentage differences (APDs) in geometric measures (length, span, volume, surface area) compared to baseline methods, indicating more accurate bundle geometry.
  • Connectomics: ATM-derived connectomes showed moderate Pearson correlations with ground truth and outperformed other methods in graph-theoretic measures.

Limitations and Future Work

  • Anatomical Detail Encoding: While ATM captures anatomical context, it may miss subtle variations in white matter structure, potentially leading to inaccuracies in complex cases. Further research is needed to understand precisely which anatomical features ATM learns.
  • Spurious Streamlines: The generative nature of ATM can sometimes lead to spurious or noisy streamlines, which could affect downstream analyses. Future work could involve optimizing model architectures and loss functions, or integrating advanced filtering techniques like FINTA.
  • Generalizability to Pathologies: ATM is currently trained on healthy brain data. Its performance may be reduced in individuals with neurological disorders or atypical white matter configurations. Future work could involve retraining ATM with data from pathological populations or incorporating conditioning mechanisms for diverse anatomical variations.
  • Multimodal Integration: The authors suggest incorporating additional imaging modalities like T2w MRI to provide a more comprehensive view of brain anatomy. They also propose adapting ATM to integrate diffusion-derived features, creating a unified framework for tractography.

Summary updated

Abstract

Diffusion tractography, a cornerstone of white matter mapping, relies on point-to-point streamline propagation, a process often limited by the signal-to-noise ratio and spatioangular resolution of diffusion MRI (dMRI). Here, we present Anatomy-to-Tract Mapping (ATM), a framework that generates bundle-specific streamlines directly from T1-weighted MRI without requiring orientation field estimation, voxelwise segmentation, or streamline propagation. ATM leverages the high quality and minimal distortion of anatomical MRI and learns from paired T1w and tractogram data to synthesize anatomically plausible, subject-specific streamlines. This anatomy-driven approach addresses complex configurations such as crossing, kissing, and bending fibers, providing robust bundle reconstructions. Using the TractoInferno dataset with 30 white matter bundles, we evaluate ATM against diffusion-based methods, including MRtrix probabilistic tracking with BundleSeg and SCIL atlas warping. ATM demonstrates strong performance across multiple metrics, including bundle similarity, volume coverage, angular correlation, and geometric fidelity.
BibTeX
@inproceedings{tan2025anatomy,
  author = {Yee Fan Tan and Khoi Minh Huynh and Siyuan Liu and Raphaël C.‐W. Phan and Chee‐Ming Ting and Pew‐Thian Yap},
  title = {{Anatomy-to-tract mapping infers white matter pathways without diffusion streamline propagation}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-66615-w},
}
Figure