Segmenting thalamic nuclei from manifold projections of multi-contrast MRI

Chang Yan, Muhan Shao, Zhangxing Bian, Anqi Feng, Yuan Xue, Jiachen Zhuo, Rao P. Gullapalli, Aaron Carass, Jerry L. Prince

TL;DR

Researchers have developed a new way to map the tiny, hard-to-see structures within the brain's thalamus by combining multiple types of MRI scans with UMAP dimensionality reduction. By projecting complex, high-dimensional imaging data into a simpler "latent space," they can accurately parcellate thalamic nuclei using a simple k-nearest neighbor approach, achieving results comparable to state-of-the-art deep learning methods.

What Problem Does This Paper Address?

The thalamus is the brain's central "relay station," processing almost all sensory and motor information. It isn't just one uniform lump of gray matter; it’s composed of several distinct clusters called nuclei. Changes in these nuclei are linked to conditions like Parkinson’s disease, multiple sclerosis, and traumatic brain injury (mTBI).

The challenge is that these nuclei are incredibly difficult to see on standard MRI scans because they have very low contrast relative to one another. While researchers have tried using diffusion imaging or specialized sequences, most existing methods either look at too few nuclei or fail to leverage the full "buffet" of available MRI contrasts. This paper seeks a way to integrate diverse imaging data to create a reliable, automated map of these sub-structures.

Methods

The authors propose a pipeline that treats MRI voxels not just as single intensity values, but as high-dimensional feature vectors.

  1. Feature Extraction: For every voxel in the thalamus, they built a massive feature vector (up to 161 dimensions). This included standard T1 and T2 weights, diffusion metrics (like FA and mean diffusivity), fiber orientation coefficients, and "synthetic multi-TI" images (simulating different inversion times). They also included 3D spatial coordinates and connectivity maps derived from tractography.
  2. Dimensionality Reduction (UMAP): To make sense of this high-dimensional data, they used Uniform Manifold Approximation and Projection (UMAP). UMAP finds the underlying "shape" of the data and projects it into a low-dimensional (2D, 3D, or 4D) latent space.
  3. Classification: Once the data is in this simplified latent space, the nuclei naturally cluster together. The researchers used a simple k-Nearest Neighbor (k-NN) algorithm to assign labels to new, unseen voxels based on their proximity to manually labeled examples in that latent space.

Datasets

The study used data from 44 subjects involved in a study of mild traumatic brain injury (mTBI).

  • Imaging: Included MPRAGE (T1-w), 3D T2-weighted, DTI, and FGATIR sequences.
  • Ground Truth: 36 subjects were semi-manually labeled. The researchers started with the Morel atlas (a standard anatomical reference), registered it to the scans, and then manually refined the labels (reducing 19 labels to 13 per hemisphere) to correct for registration errors.
  • Validation: 30 viable subjects were used in a five-fold cross-validation.

Key Insights and Ideas

  • The "Curse of Dimensionality" vs. Connectivity: Surprisingly, adding cortical connectivity data (information about where brain fibers lead) actually decreased accuracy slightly. The authors suggest that the extra dimensions might have introduced more noise than signal, or that the manual delineators didn't use connectivity cues, so the model couldn't learn from them.
  • Natural Clustering: The core "aha" moment is that thalamic nuclei are naturally separated in high-dimensional feature space. You don't necessarily need a complex neural network to find them; you just need a way to project that manifold down to a space where a simple distance-based classifier can work.
  • Spatial Context Matters: The best results came from a 63-dimension vector that combined "Base" MRI features, spatial coordinates, and multi-TI synthetic images. This suggests that where a voxel is located is just as important as what its intensity is.

Key Results

  • Accuracy: The method achieved a volume-weighted Dice score of approximately 0.64 to 0.71 (depending on the latent space dimension). This is considered comparable to current state-of-the-art automated parcellation methods.
  • Latent Space Trade-offs: Moving from a 2D to a 4D latent space improved accuracy (especially for smaller nuclei), but at a significant computational cost. A 2D embedding took ~10 minutes to train, while a 4D embedding took ~2 hours.
  • Specific Nuclei: Some nuclei, like the Mediodorsal (MD) and Inferior Pulvinar (PuI), were segmented with very high accuracy (Dice > 0.80), while smaller, more obscure nuclei remained a challenge.

Limitations and Future Work

  • Sample Size: The study relied on a relatively small group of 30 subjects. Expanding this to a larger, more diverse population would help prove the model's robustness.
  • Manual Labeling Bias: The "ground truth" was created by humans looking at specific MRI contrasts. If the humans couldn't see a boundary, the model likely couldn't learn it either.
  • Computational Efficiency: While UMAP is faster than older techniques like t-SNE, the jump to higher-dimensional latent spaces (4D+) currently hits a wall of diminishing returns regarding time vs. accuracy. Future work could involve optimizing this embedding process.

Summary updated

Abstract

The thalamus is a subcortical gray matter structure that plays a key role in relaying sensory and motor signals within the brain. Its nuclei can atrophy or otherwise be affected by neurological disease and injuries including mild traumatic brain injury. Segmenting both the thalamus and its nuclei is challenging because of the relatively low contrast within and around the thalamus in conventional magnetic resonance (MR) images. This paper explores imaging features to determine key tissue signatures that naturally cluster, from which we can parcellate thalamic nuclei. Tissue contrasts include T1-weighted and T2-weighted images, MR diffusion measurements including FA, mean diffusivity, Knutsson coefficients that represent fiber orientation, and synthetic multi-TI images derived from FGATIR and T1-weighted images. After registration of these contrasts and isolation of the thalamus, we use the uniform manifold approximation and projection (UMAP) method for dimensionality reduction to produce a low-dimensional representation of the data within the thalamus. Manual labeling of the thalamus provides labels for our UMAP embedding from which k nearest neighbors can be used to label new unseen voxels in that same UMAP embedding. N-fold cross-validation of the method reveals comparable performance to state-of-the-art methods for thalamic parcellation.
BibTeX
@inproceedings{yan2023segmenting,
  author = {Chang Yan and Muhan Shao and Zhangxing Bian and Anqi Feng and Yuan Xue and Jiachen Zhuo and Rao P. Gullapalli and Aaron Carass and Jerry L. Prince},
  title = {{Segmenting thalamic nuclei from manifold projections of multi-contrast MRI}},
  booktitle = {Image Processing},
  year = {2023},
  doi = {10.1117/12.2654405},
}

Loading related content...

Figure