A Contrast-Agnostic Method for Ultra-High Resolution Claustrum Segmentation

TL;DR

This work addresses the challenge of automatic claustrum segmentation despite its thin sheet and variable MRI contrast by proposing a contrast- and resolution-agnostic framework built on SynthSeg that learns from ultra-high-resolution ex vivo labels and generalizes to typical in vivo MRI. The approach uses synthetic training intensities and label augmentations to train a 3D U-Net, coupled with a quality-control procedure based on nonlinear registration to MNI152, enabling robust segmentation across T1-, T2-, and other contrasts. The authors provide 18 high-resolution manual labels, demonstrate cross-dataset generalization (IXI, Miriad, FSM), show reasonable Dice scores with strong robustness to downsampling and modality changes, and release code and data resources. The method yields practical benefits for studying claustrum structure and function, offering a versatile tool for high- and standard-resolution MRI and potential post hoc corrections for surrounding regions.

Abstract

The claustrum is a band-like gray matter structure located between putamen and insula whose exact functions are still actively researched. Its sheet-like structure makes it barely visible in in vivo Magnetic Resonance Imaging (MRI) scans at typical resolutions and neuroimaging tools for its study, including methods for automatic segmentation, are currently very limited. In this paper, we propose a contrast- and resolution-agnostic method for claustrum segmentation at ultra-high resolution (0.35 mm isotropic); the method is based on the SynthSeg segmentation framework (Billot et al., 2023), which leverages the use of synthetic training intensity images to achieve excellent generalization. In particular, SynthSeg requires only label maps to be trained, since corresponding intensity images are synthesized on the fly with random contrast and resolution. We trained a deep learning network for automatic claustrum segmentation, using claustrum manual labels obtained from 18 ultra-high resolution MRI scans (mostly ex vivo). We demonstrated the method to work on these 18 high resolution cases (Dice score = 0.632, mean surface distance = 0.458 mm, and volumetric similarity = 0.867 using 6-fold Cross Validation (CV)), and also on in vivo T1-weighted MRI scans at typical resolutions (~1 mm isotropic). We also demonstrated that the method is robust in a test-retest setting and when applied to multimodal imaging (T2-weighted, Proton Density and quantitative T1 scans). To the best of our knowledge this is the first accurate method for automatic ultra-high resolution claustrum segmentation, which is robust against changes in contrast and resolution. The method is released at https://github.com/chiara-mauri/claustrum_segmentation and as part of the neuroimaging package Freesurfer (Fischl, 2012).

Figures (15)

• Figure 1: Example of an ex vivo hemisphere (case 14, left hemisphere, axial view, slice 789, voxel size 0.12 mm), where a cropped region around claustrum is highlighted. Note that the contrast between claustrum and white matter faints in the anterior end of claustrum (blue arrow), raising challenges in the delineation of the structure.
• Figure 2: (a-b-c): Different views of an ex vivo hemisphere (case 14, left hemisphere) cropped around claustrum, with overimposed manual label. The coronal view (b) shows both dorsal and ventral portions of the claustrum, and was used for manual tracing. The dorsal part is thinner, enclosed between the external (A) and the extreme (B) capsules, and ultimately wraps around the superior cortical gyri (C). The ventral portion is wider, displays the characteristic "fingers" (D), and closely follows the contour of the ventral amygdala (E). (d-e): 3D rendering of the claustrum manual label of the same case, shown from two different angles. Note that the structure has only one connected component and that the fragmented look of the ventral "fingers" appears only in 2D views.
• Figure 3: Two coronal views of an ev vivo hemisphere (case 14, left hemisphere) cropped around claustrum, with overimposed manual label. Light-blue arrows point to ventral "fingers" that have been annotated individually, while red arrows highlight regions where they were labeled as a whole, due to voxel size limitations.
• Figure 4: Overview of the pipeline used to create training labels. Case 14 (left hemisphere) is shown. In the merged figure, white arrows highlight regions where whole-brain SynthSeg incorrectly labeled the claustrum as putamen. The claustrum is shown with partial transparency to make these errors visible. In the subsequent figure, yellow arrows indicate voxels that were manually reassigned to white matter (WM); these appear as black regions in the image.
• Figure 5: Example of axial (left, slice number 270), coronal (middle, slice number 247), and sagittal (right, slice number 118) view of a label map (case 14, left hemisphere flipped on the right side), obtained by overimposing the claustrum manual label to the automatic segmentation of surrounding structures, which were used to train the SynthSeg segmentation model. The claustrum label is at 0.35 mm resolution, whereas the other structures are output by the whole-brain SynthSeg at 1 mm. The slices shown here correspond to those in Fig. \ref{['main:fig:manual_label']}, given the claustrum voxel sizes of 0.35 mm vs 0.12 mm respectively.