RATNUS: Rapid, Automatic Thalamic Nuclei Segmentation Using Multimodal MRI Inputs

Anqi Feng, Zhangxing Bian, Blake E. Dewey, Alexa Gail Colinco, Jiachen Zhuo, Jerry L. Prince

Abstract

Accurate segmentation of thalamic nuclei is important for better understanding brain function and improving disease treatment. Traditional segmentation methods often rely on a single T1-weighted image, which has limited contrast in the thalamus. In this work, we introduce RATNUS, which uses synthetic T1-weighted images with many inversion times along with diffusion-derived features to enhance the visibility of nuclei within the thalamus. Using these features, a convolutional neural network is used to segment 13 thalamic nuclei. For comparison with other methods, we introduce a unified nuclei labeling scheme. Our results demonstrate an 87.19% average true positive rate (TPR) against manual labeling. In comparison, FreeSurfer and THOMAS achieve TPRs of 64.25% and 57.64%, respectively, demonstrating the superiority of RATNUS in thalamic nuclei segmentation.

Introduction

The thalamus serves as a critical relay station in the brain, composed of numerous small, functionally distinct clusters called nuclei. These structures are implicated in various neurological conditions, including Parkinson’s disease, Alzheimer’s disease, and Multiple Sclerosis. However, standard 3T MRI scans, such as T1-weighted MPRAGE, often lack the necessary contrast to distinguish these nuclei, leading many researchers to treat the thalamus as a single mass. While specialized tools exist, they frequently require high-field 7T magnets or suffer from inconsistent naming conventions that hinder cross-study comparisons.

Methodology

The authors propose RATNUS (Rapid, Automatic Thalamic Nuclei Segmentation), a deep-learning framework utilizing a two-step 3D U-Net architecture. The model processes a 68-channel input designed to maximize structural contrast:

  1. Synthetic Multi-TI Images: Rather than using a single T1 scan, the system calculates Proton Density (PD) and T1 maps from MPRAGE and FGATIR sequences. It then synthesizes a series of 51 images with varying inversion times (TI), effectively "sweeping" through the T1 recovery curve to highlight different nuclei boundaries.
  2. Diffusion-Derived Features: The model incorporates fractional anisotropy (FA) and 5D Knutsson mapping. These features represent white matter tract orientations, providing critical boundary information where traditional intensity-based contrast fails.
  3. Sparse Label Training: To account for the difficulty of manual labeling, the network was trained using a masked Dice Loss function. This allows the model to learn exclusively from high-confidence, manually labeled voxels while ignoring unlabeled regions.
Multimodal MRI inputs and synthetic multi-TI image generation

Datasets and Evaluation

The study utilized a dataset of 24 subjects, including healthy controls and patients with mild traumatic brain injury. To facilitate benchmarking, the authors developed a Labeling Schemes Unifying Protocol, which maps the disparate outputs of tools like FreeSurfer and THOMAS into seven major categories based on the Morel Atlas.

Results and Discussion

RATNUS demonstrated significant improvements over existing industry standards. It achieved an average True Positive Rate (TPR) of 87.19% against manual labels. When evaluated using the unified protocol, RATNUS maintained a TPR of 75.80%, whereas FreeSurfer (64.25%) and THOMAS (57.64%) showed lower performance. Qualitative analysis indicated that RATNUS produces cleaner segmentations that adhere more closely to anatomical boundaries, avoiding the over-segmentation common in FreeSurfer or the under-segmentation seen in THOMAS.

Thalamic nuclei segmentation map and corresponding manual label color map

Conclusion and Future Work

By leveraging the full spectrum of T1 contrast and diffusion orientation, RATNUS provides a robust solution for thalamic segmentation on common 3T MRI hardware. Future work will focus on removing the dependency on FGATIR sequences by synthesizing the necessary multi-TI images from standard clinical scans and expanding the training set through semi-supervised learning to improve generalization across diverse populations.

Summary updated

BibTeX
@inproceedings{feng2024ratnus,
  author = {Anqi Feng and Zhangxing Bian and Blake E. Dewey and Alexa Gail Colinco and Jiachen Zhuo and Jerry L. Prince},
  title = {{RATNUS: Rapid, Automatic Thalamic Nuclei Segmentation Using Multimodal MRI Inputs}},
  booktitle = {LDTM/MMMI/ML4MHD/ML-CDS @ MICCAI},
  year = {2024},
  doi = {10.1007/978-3-031-84525-3\_14},
}
Figure