CATNUS: Coordinate-Aware Thalamic Nuclei Segmentation Using T1-Weighted MRI

TL;DR

CATNUS addresses automatic segmentation of 13 thalamic nuclei from MRI using a coordinate-aware 3D U-Net that incorporates voxel coordinates to improve localization, especially for small nuclei. It supports inputs from quantitative T1 maps, MPRAGE, or FGATIR via a unified protocol and outperforms established tools in segmentation accuracy and test-retest reliability, with strong generalization to traveling-subject datasets across scanners and vendors, particularly after intensity harmonization. The approach combines multi-TI derived contrasts for labeling with a sparse but high-quality manual annotation protocol and post-processing that enforces anatomical plausibility. Overall, CATNUS offers a fast, accurate, and generalizable solution for thalamic nuclei segmentation, suitable for large-scale neuroimaging studies and clinical assessments, with pretrained models and open science availability planned.

Abstract

Accurate segmentation of thalamic nuclei from magnetic resonance images is important due to the distinct roles of these nuclei in overall brain function and to their differential involvement in neurological and psychiatric disorders. However, segmentation remains challenging given the small size of many nuclei, limited intrathalamic contrast and image resolution, and inter-subject anatomical variability. In this work, we present CATNUS (Coordinate-Aware Thalamic Nuclei Segmentation), segmenting 13 thalamic nuclei (or nuclear groups) using a 3D U-Net architecture enhanced with coordinate convolution layers, which provide more precise localization of both large and small nuclei. To support broad clinical applicability, we provide pre-trained model variants that can operate on quantitative T1 maps as well as on widely used magnetization-prepared rapid gradient echo (MPRAGE) and fast gray matter acquisition T1 inversion recovery (FGATIR) sequences. We benchmarked CATNUS against established methods, including FreeSurfer, THOMAS and HIPS-THOMAS, demonstrating improved segmentation accuracy and robust test-retest reliability across multiple nuclei. Furthermore, CATNUS demonstrated strong out-of-distribution generalization on traveling-subject datasets spanning multiple scanners, field strengths, and vendors, producing reliable and anatomically coherent segmentations across diverse acquisition conditions. Overall, CATNUS provides an accurate and generalizable solution for thalamic nuclei segmentation, with strong potential to facilitate large-scale neuroimaging studies and support real-world clinical assessment.

Figures (10)

• Figure 1: Overview of the CATNUS preprocessing and annotation workflow. A. Data preprocessing and computation pipeline. MPRAGE and FGATIR images are co-registered to the MNI152 space, followed by joint N4 bias field correction and white matter (WM) mean normalization. Quantitative T1 and PD maps are computed, and multi-TI images are generated using the estimated T1 and PD maps. B. Visualization of thalamic structures on different images. Representative MPRAGE, FGATIR, T1 and PD maps, and multi-TI images illustrate how varying inversion times enhance intra-thalamic visibility and facilitate nuclear delineation. C. Manual label delineation. Summary of the 13 thalamic nuclei with their abbreviations, full names, selected TI values used for annotation, and the color coding scheme. D. Sparse-labeling problem and solution. Annotations were restricted to high-confidence voxels (left). Unlabeled voxels were explicitly separated from background and excluded from model training (right). E. Selected multi-TI images for annotation. Examples of nuclei annotated on specific TI images, illustrating how different inversion times highlight distinct nuclear boundaries. Red and yellow outlines indicate the thalamic boundary and the corresponding nucleus boundary, respectively. When only yellow is visible, the nucleus boundary coincides with the thalamic boundary on this slice.
• Figure 2: Overview of the CATNUS segmentation framework. A. Training Phase. Input images undergo spatial and intensity augmentations, followed by center-cropping to a fixed size. The model is supervised using Dice loss computed only on voxels with manual labels, while unlabeled voxels are masked out from loss computation. B. Inference Phase. A cropped image is passed through the model, and each voxel is assigned the class with the highest predicted probability. The output is then padded back to the original volume size. C. Post Processing. Predicted segmentations are refined sequentially, first by thalamus-level filtering, then by nucleus-level refinement. D. Segmentation Model. The model takes an image and its voxel-wise spatial coordinates (x, y, z) as input, processes them via coordinate convolution, and outputs probability maps using a 4-level 3D U-Net.
• Figure 3: A comprehensive mapping of thalamic nuclei categories between CATNUS and the definitions from FreeSurfer, THOMAS and HIPS-THOMAS. The first column shows the major categories derived from our unified grouping strategy, each corresponding to merged nuclei listed in the second, third, and fourth columns. All major categories and individual nuclei are assigned distinct colors to help with visual differentiation. Note that THOMAS and HIPS-THOMAS share the same labeling and color conventions. Abbreviations for nuclei in FreeSurfer, THOMAS, and HIPS-THOMAS are adopted from iglesias2018probabilistic, su2019thalamus and vidal2024robust.
• Figure 4: Within-domain segmentation performance across three input modalities (T1 map, MPRAGE, and FGATIR) with or without coordinate convolution (CoordConv). Each boxplot shows the distribution of true positive rates (TPR) for one nucleus or the volume-weighted average (VWA), grouped by modality and CoordConv configuration. Performance was evaluated across 24 manually labeled subjects using 8-fold cross-validation. Statistically significant differences relative to T1map (Coord) are indicated by stars (*p<0.05, **p<0.01, ***p<0.001). The top-right legend indicates color mapping for the six configurations. The horizontal bar summarizes the volume proportions of each nucleus.
• Figure 5: Qualitative comparisons of segmentation results across input modalities (T1 map, MPRAGE, FGATIR) and coordinate convolution (CoordConv) settings, overlaid on FGATIR images. Each row shows a representative slice in a different anatomical plane (axial, sagittal, coronal), with each slice taken from a different subject. Ground-truth annotations (leftmost column) are compared against model predictions from the six configurations. Yellow circles highlight common failure cases, especially for the small PuA nucleus, which is frequently missed when CoordConv is not used. Pink arrows indicate additional mislocalizations of the PuA nucleus along thalamic boundaries.