A comprehensive protocol for manual segmentation of the human claustrum and its sub-regions using high-resolution MRI

The claustrum (Cl) is a thin grey matter structure located in the center of each brain hemisphere. Cl has been hypothesized as a central hub of the brain for multisensory/sensorimotor integration, consciousness, and attention. Accumulating evidence has suggested that Cl might be important in the development of severe neurological and psychiatric symptoms including epileptic seizures and psychosis. However, the specifics of the roles of Cl in human epilepsy and psychosis are largely unknown, primarily due to methodological limitations related to the thin morphology of Cl that is challenging to delineate accurately using conventional methods. The goal of this work is to develop noninvasive multimodal neuroimaging methods to delineate Cl anatomy by utilizing a large healthy adult high resolution (0.7mm3) T1-weighted MRI collected as part of the Washington University-Minnesota Consortium Human Connectome Project (WU-Minn HCP). We developed a comprehensive manual segmentation protocol to delineate Cl based on a cellular level brain atlas. The protocol involves detailed guidelines to delineate the three subregions of Cl, including the dorsal, ventral, and temporal Cl that can be parcellated based on a geometric method. As demonstrated in a representative result, Cl is large in its anterior-posterior, and the dorsal-ventral extent. Also, the volume is comparable to that of the amygdala. It is required to assess the reliability of the protocol so that it can be used for future anatomical studies of neuropsychiatric disorders, including epilepsy and schizophrenia.

Annotations for: A comprehensive protocol for manual segmentation of the human claustrum and its sub-regions using high-resolution MRI

✅ Key Points

The claustrum is a thin deep-brain structure located in the basolateral telencephalon of the mammalian brain [@kangComprehensiveProtocolManual2020, p. 1]

Dataset [@kangComprehensiveProtocolManual2020, p. 2]

from the WUMinn HCP database [@kangComprehensiveProtocolManual2020, p. 2]

Structural scan [@kangComprehensiveProtocolManual2020, p. 2]

Structural images were acquired using the 3D MPRAGE [@kangComprehensiveProtocolManual2020, p. 2]

Used 3D slicer to do the manual claustrum labelling [@kangComprehensiveProtocolManual2020, p. 2]

3D Slicer software [@kangComprehensiveProtocolManual2020, p. 2]

Important point to note is that the claustrum is best to label using the axial view for the dorsal claustrum and the ventral claustrum to be labelled in the ventral view [@kangComprehensiveProtocolManual2020, p. 3]

Relying on a single-plane perspective to trace the entire claustrum can easily result in errors, such as the inclusion of adjacent cortical and subcortical voxels from transitional areas that appear connected to the claustrum. [@kangComprehensiveProtocolManual2020, p. 3]

Isotopic voxel sizes? [@kangComprehensiveProtocolManual2020, p. 4]

Despite the chal-lenging anatomical features of the anterior temporal claustrum, high-resolution MRIs with a 0.7mm3 isotropic voxel areas [@kangComprehensiveProtocolManual2020, p. 4]

Redundant [@kangComprehensiveProtocolManual2020, p. 4]

high-resolution MRIs with a 0.7mm3 isotropic voxel size [@kangComprehensiveProtocolManual2020, p. 4]

How can you be sure that the FreeSurfer segmentation correctly labelled the putamen? Sometimes even at 0.7mm3 the freesurfer segmentation fails and can potentially leak into the claustrum laterally [@kangComprehensiveProtocolManual2020, p. 6]

m Matlab program automatically identifies the two landmark points and the anterior-posterior limits following the same anatomical criteria as the manual method and using FreeSurfer labels of the putamen and amyg-dala. [@kangComprehensiveProtocolManual2020, p. 6]

ICV-adjusted, relative volumes account for individual differences in head size, enabling more accurate comparisons across subjects by normalizing regional volumes relative to overall brain size. [@kangComprehensiveProtocolManual2020, p. 6]

*So JB labelled the claustrum twice? This is not clear here. I would rather clarify to say that 2 independent operators labelled the claustrum followed by an additional re-tracing of the claustrum by JB. * [@kangComprehensiveProtocolManual2020, p. 6] ** Are all 3 of the tracing then used in the DICE coefficient estimates, even though JB labelled the claustrum twice would this not inflate the DICE estimate if JB labelled the claustrum in the same way two times?

Two operators (JB and KM) independently traced the entirety of the whole claustrum using the protocol described above to assess the inter-rater reliability. Additionally, to assess intrarater reliability, one of the operators (JB) retraced the claus-trum. [@kangComprehensiveProtocolManual2020, p. 6]

Again, here the same rater did the identifying of landmarks, would this not inflate the DICE score? [@kangComprehensiveProtocolManual2020, p. 6]

The third operator identified the landmark points twice to assess in-tra-rater reliability. [@kangComprehensiveProtocolManual2020, p. 6]

We assessed the reliability of the subregion parcellation protocol using the first whole claustrum tracing sets from the first operator, which allows controlling for the confounding effect of inter-rater variation in the whole claustrum tracings (Entis et al., 2012). [@kangComprehensiveProtocolManual2020, p. 6]

The intraclass correlation. Need to read a bit more about what exactly this is? [@kangComprehensiveProtocolManual2020, p. 6]

The ICC, specifically a two-way mixed-effects model with absolute agreement, quantifies the consistency and agreement of volumetric measurements between operators, accounting for systematic differences. [@kangComprehensiveProtocolManual2020, p. 6]

Sad [@kangComprehensiveProtocolManual2020, p. 6]

MATLAB version 2023a [@kangComprehensiveProtocolManual2020, p. 6]

Would love to get my hands on the 3D models [@kangComprehensiveProtocolManual2020, p. 6]

3D models illustrate the unique shape of the claustrum (Figure 5.C). [@kangComprehensiveProtocolManual2020, p. 6]

Why not? [@kangComprehensiveProtocolManual2020, p. 8]

, our in vivo MRI-based volumetric meas-urement may not be directly comparable to the postmortem brain studies. [@kangComprehensiveProtocolManual2020, p. 8]

This is true [@kangComprehensiveProtocolManual2020, p. 8]

given the lack of detailed in-formation on the manual segmentation protocol in the older studies, differences in the segmentation method may have con-tributed to the differences in the claustrum volume, especially in the substantially lower volumes reported from Bernstein et al. (2016b). [@kangComprehensiveProtocolManual2020, p. 8]

Zald and Pardo 2002 used a 10mm FWHM Gaussian smoothing filter therefore not the most representative study here! [@kangComprehensiveProtocolManual2020, p. 9]

For example, con-sistent with the hypothesis that Cl plays a critical role in sali-ence processing (Remedios et al., 2010; Reser et al., 2014b; Smith et al., 2019), a human fMRI study found activations in the periamygdala region including tCl in response to emotion-ally salient (aversive) auditory stimuli (Zald & Pardo, 2002) [@kangComprehensiveProtocolManual2020, p. 9]

However, the success of automated segmentation tools largely depends on the quality of MRI scans and the availability of accurate and reliable claustrum segmentations, which serve as the ground truth for deep-learning algorithm development. [@kangComprehensiveProtocolManual2020, p. 9]

Either misplaced or redundant [@kangComprehensiveProtocolManual2020, p. 9]

Ad-ditionally, the present manual segmentation protocol will serve multiple crucial purposes in neuroimaging research. [@kangComprehensiveProtocolManual2020, p. 9]

Or in other words 0.1mm3 ? [@kangComprehensiveProtocolManual2020, p. 9]

< 0.53 mm voxel size; Coates & Zaretskaya, 2024 [@kangComprehensiveProtocolManual2020, p. 9]

Spell check. Brains [@kangComprehensiveProtocolManual2020, p. 9]

adapt the protocol for pediatric brain [@kangComprehensiveProtocolManual2020, p. 9]

Yes this is true FreeSurfer segmentation may have some errors and as a result result in a poor automatic claustrum segmentation. Have the authors not considered the use of the Nextbrain atlas as part of the newer FreeSurfer versions that may improve the reliability of segmenting areas around the claustrum? [@kangComprehensiveProtocolManual2020, p. 9]

Finally, since the fully automatic parcellation of claustral subregions relies on FreeSurfer’s segmentation, its accuracy may be compromised by FreeSurfer-generated errors. In our study, five subjects exhibited minor errors, such as mislabeling of a small number of anterior-inferior temporal claustrum voxels extending beyond the amygdala’s anterior boundary. These errors were minimal and did not significantly affect reli-ability. However, final visual inspection and correction, as in the semi-automatic approach, are recommended to enhance ac-curacy. [@kangComprehensiveProtocolManual2020, p. 9]

Background

Located at the center of each hemisphere, it is the brain’s most highly connected hub [@kangComprehensiveProtocolManual2020, p. 1]

Hypothesis

The claustrum has dense anatomical connections with cortical areas, including temporal, motor, somatosensory, visual, auditory, limbic, associative, sensorimotor, and prefrontal cortices (Brown et al., 2017; Goll et al., 2015; Milardi et al., 2015; Rodríguez-Vidal et al., 2024). [@kangComprehensiveProtocolManual2020, p. 1]

Here the authors mention that the inflated high reliability might also actually come from using the same person to trace the claustrum twice [@kangComprehensiveProtocolManual2020, p. 8]

To control for the confounding effect of inter-rater variability in whole-claustrum tracing on the subregional parcellation, we used the same whole-claustrum tracing set from a single operator for reliability assessment. While this factor may have contributed to the high reliability, more importantly, our method employs a geometric approach that identifies two landmark points based on the clearly discernible subcortical structure (i.e., the putamen) and an automatic process that parcellates the subregions based on these landmark points. [@kangComprehensiveProtocolManual2020, p. 8]

Methods

Next brain atlas is a new probabilistic atlas that does include a claustrum label. How does this manual claustrum segmentation approach compare with this probabilistic atlas? The authors do not take into account that automatic claustrum segmentation approaches have since been improved since their first preprint in 2020. [@kangComprehensiveProtocolManual2020, p. 2]

However, widely used probabilistic brain atlases (e.g., Harvard-Oxford cortical and subcortical brain atlas) and a recent brain atlas based on connectional architecture (e.g., Brainnetome atlas; http://atlas.brainnetome.org/index.html) do not include the claustrum. [@kangComprehensiveProtocolManual2020, p. 2]

FreeSurfer 8.0.0 now comes shipped with the Next brain atlas and so does provide automated parcellation in the sense of a probabilistic atlas. What do the authors mean here exactly? [@kangComprehensiveProtocolManual2020, p. 2]

no neuroimaging software provides automated parcellation [@kangComprehensiveProtocolManual2020, p. 2]

Okay so assessing the reliability of the manual claustrum segmentation approach based on inter and intra-rater reliability is only a partial picture. I think the authors may want to also show reliability or even validity that their claustrum manual claustrum segmentation method shows a marked improvement over other such methods, particularly the new nextbrain claustrum atlas [@kangComprehensiveProtocolManual2020, p. 2]

assessed the reliability of the man-ual segmentation protocol by examining intra- and inter-rater reliability. T [@kangComprehensiveProtocolManual2020, p. 2]

B in iii is not visible [@kangComprehensiveProtocolManual2020, p. 5]

Figure 4. [@kangComprehensiveProtocolManual2020, p. 5]

Doesn’t make sense [@kangComprehensiveProtocolManual2020, p. 6]

Then the program Step-2 p [@kangComprehensiveProtocolManual2020, p. 6]

The consistently high reliability across all subregions validates the robustness of our parcellation protocol [@kangComprehensiveProtocolManual2020, p. 7]

Remedios et al 2014 is more about salience the 2010 paper specifically examined the sensory specific zones of the claustrum [@kangComprehensiveProtocolManual2020, p. 9]

For example, con-sistent with the hypothesis that Cl plays a critical role in sali-ence processing (Remedios et al., 2010 [@kangComprehensiveProtocolManual2020, p. 9]

Not sure if this is completely relevant to the current paper. Coates et al 2024 determined the extent of sensory specific activation found in macaques applies to humans similar to remedios et al 2010 [@kangComprehensiveProtocolManual2020, p. 9]

As a growing body of functional neuroimaging studies have revealed additional roles of the claustrum (Coates et al., 2024; Rodríguez-Vidal et al., 2024; Stewart et al., 2024), future stud-ies should systematically investigate its functions using various experimental paradigms and large samples. [@kangComprehensiveProtocolManual2020, p. 9]

Results

? Why is this sentence here? What’s the purpose of it? I don’t quite get what the point here is? Yeah so the claustrum volume is quite large but then why would this be a counterpoint to poor segmentation as a result of no claustrum segmentation protocol method not being used [@kangComprehensiveProtocolManual2020, p. 8]

it is notable that the claustral volumes are comparable to the amygdala volume manually seg-mented from high-resolution MRI (left: 1727.6±177.9 mm³; right: 1750.3±218.9 mm³; Entis et al., 2012), indicating that the claustrum is a sizeable subcortical structure [@kangComprehensiveProtocolManual2020, p. 8]