An efficient deep learning framework for consistent superficial white matter tractography parcellation

Tengfei Xue, Fan Zhang, Chaoyi Zhang, Yuqian Chen, Yang Song, Nikos Makris, Yogesh Rathi, Weidong Cai, Lauren J. O’DonnellProceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition 2023 · DOI

TL;DR

This paper introduces SupWMA, a novel deep learning framework for efficiently and accurately parcellating superficial white matter (SWM) tracts in the brain. By adapting a point-cloud-based network and incorporating supervised contrastive learning, SupWMA achieves state-of-the-art performance in identifying SWM clusters, outperforming existing methods in both accuracy and speed.

What Problem Does This Paper Address?

Diffusion MRI (dMRI) tractography allows us to map the brain's white matter connections non-invasively. However, raw tractography data, consisting of millions of "streamlines," is difficult for clinicians and researchers to use directly for analysis or visualization. Tractography parcellation aims to organize these streamlines into meaningful bundles or clusters. While significant progress has been made in parcellating deep white matter (DWM) tracts, the superficial white matter (SWM) – which comprises shorter, more complex connections between adjacent cortical areas (often called "u-fibers") – remains a challenge. Existing SWM parcellation methods often rely on manual region-of-interest definitions, which can be labor-intensive and subject to variability, or use clustering techniques that struggle with consistency across subjects and can be computationally expensive.

Methods

The core of SupWMA is a deep learning framework built upon a point-cloud-based network. Unlike traditional methods that might represent streamlines as images, SupWMA treats each streamline as a set of 3D points. This approach is advantageous because it inherently preserves the spatial information of the streamline and is robust to the directionality of the points (i.e., a streamline traced from cortex to subcortical area or vice versa will have the same representation).

Key methodological innovations include:

  • Modified PointNet Architecture: The authors adapt the PointNet architecture, a standard for processing point clouds, by removing its "transformation nets." This modification is crucial for preserving the absolute spatial location of streamlines within the brain, which is vital for accurate parcellation.
  • Supervised Contrastive Learning (SCL): To improve the network's ability to distinguish between genuine SWM tracts and spurious or outlier streamlines (which can have similar geometric shapes), SupWMA employs SCL. This technique trains the network to pull representations of streamlines belonging to the same class closer together in a latent space while pushing apart representations of streamlines from different classes. This results in more discriminative features.
  • Two-Phase Training: The framework is trained in two phases: first, using SCL to learn robust feature representations, and second, using a standard cross-entropy loss for the final classification of streamlines into SWM clusters and a "non-SWM cluster" (which includes DWM tracts and outliers).

Datasets

  • Training and Validation: A large, anatomically curated dataset of 1 million labeled streamlines was used. This dataset was derived from 100 healthy young adults and contained labels for 198 expert-curated SWM clusters, with all other streamlines (DWM and outliers) grouped into a single "non-SWM cluster."
  • Testing: Three independently acquired, publicly available datasets were used for evaluation, comprising a total of 150 subjects. These datasets represent diverse populations in terms of age (children, teenagers, adults, older adults) and health conditions (healthy controls, Parkinson's disease patients), allowing for assessment of the model's generalization capabilities. The datasets included:
  • Human Connectome Project (HCP)
    • Adolescent Brain Cognitive Development (ABCD)
    • Parkinson's Progression Markers Initiative (PPMI)

Key Insights and Ideas

  • Point Clouds for Streamlines: Treating streamlines as point clouds, rather than images, is a powerful way to capture their geometric and spatial properties without being sensitive to directionality. This is a significant departure from many prior deep learning approaches in tractography.
  • SCL for Outlier Rejection: The use of supervised contrastive learning is critical for effectively separating true SWM tracts from noise and outliers, which is a common problem in tractography. It helps the model learn subtle but important differences in streamline geometry and spatial context.
  • Generalization Across Diverse Datasets: SupWMA demonstrates strong performance not only on a curated ground truth dataset but also on independent datasets with varying imaging protocols, ages, and health conditions, highlighting its robustness and generalizability.
  • Efficiency is Key: The framework achieves significantly faster computation times compared to existing state-of-the-art methods, making it more practical for clinical and research applications.

Key Results

  • Accuracy: SupWMA achieved the highest accuracy and macro F1-score on the ground truth dataset, outperforming other state-of-the-art methods like DeepWMA and DCNN+CL+ATT.
  • Consistency: The framework demonstrated highly consistent parcellation results across multiple testing datasets.
  • Speed: SupWMA is significantly faster than existing methods, with computation times orders of magnitude lower. For example, it achieved a computation time of 1min 22s on a Linux workstation using CPU+GPU for a dataset with 0.44 million streamlines, compared to 2min 25s for DeepWMA.
  • Cluster Identification: On datasets without ground truth labels, SupWMA achieved the highest cluster identification rate (CIR) across all three testing datasets (HCP, ABCD, PPMI), indicating its superior ability to identify meaningful white matter clusters.

Limitations and Future Work

  • Limitations: The paper acknowledges that while SupWMA is fast, the training process, particularly the contrastive learning phase, still requires substantial computational resources. The performance on the "non-SWM cluster" is a grouping of DWM and outliers, and future work could explore finer-grained parcellation within this group.
  • Future Work: The authors suggest that SupWMA's lightweight and efficient nature, combined with its ability to preserve spatial features, makes it a promising foundation for further research in white matter analysis. Potential future directions could involve extending the framework to other tractography tasks or exploring more advanced point-cloud architectures.

Summary updated

Abstract

We propose a deep-learning-based framework, Superficial White Matter Analysis (SupWMA), which performs an efficient and consistent parcellation of 198 SWM clusters from whole-brain tractography. A point-cloud-based network is developed for our SWM parcellation task, and supervised contrastive learning enables more discriminative representations between plausible streamlines and outliers. SupWMA obtains highly consistent and accurate SWM parcellation results on a large tractography dataset with ground truth labels and on three independently acquired testing datasets from individuals across ages and health conditions.
BibTeX
@inproceedings{xue2023efficient,
  author = {Tengfei Xue and Fan Zhang and Chaoyi Zhang and Yuqian Chen and Yang Song and Nikos Makris and Yogesh Rathi and Weidong Cai and Lauren J. O’Donnell},
  title = {{An efficient deep learning framework for consistent superficial white matter tractography parcellation}},
  booktitle = {Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition},
  year = {2023},
  doi = {10.58530/2022/3302},
}
Figure