Subcortical brain structures are often neglected in neuroimaging studies due to their small size, limited inter-regional contrast, and weak signal-to-noise ratio in functional imaging (Forstmann et al., 2016; Johansen-Berg, 2013). Yet, these small and diverse structures are prominent nodes in functional networks (Marquand et al., 2017; Ji et al., 2019), and they undergo pathological alterations already at early stages of neurodegenerative diseases (Andersen et al., 2014; Koshiyama et al., 2018). Deep brain stimulation surgery, originally performed to reduce motor symptoms in essential tremors, is now a promising therapeutic option in later stages of Parkinson’s disease and movement disorders, as well as refractory psychiatric illnesses in obsessive-compulsive disorder, anorexia, or depression (Forstmann et al., 2017; Mosley et al., 2018). Evolutionary genetics even uncovered that in modern humans, Neanderthal-inherited alleles were preferentially down-regulated in subcortical and cerebellar regions compared to other brain regions (McCoy et al., 2017), suggesting these structures to be essential in making us specifically human.

Despite their importance, these areas are particularly difficult to image. Furthermore, the size, shape, and location of these brain regions changes with development and aging (Fjell et al., 2013; Keuken et al., 2013; Yeatman et al., 2014; Herting et al., 2018). Experience-based plasticity continuously remodels myelin (Tardif et al., 2016; Hill et al., 2018; Turner, 2019), iron and other magnetic substances accumulate with age or pathology (Andersen et al., 2014; Zhang et al., 2018), both bringing changes in the MRI appearance of subcortical regions with diverse tissue characteristics (Draganski et al., 2011; Keuken et al., 2017).

Thus, mapping the structure and function of the subcortex is a major endeavor as well as a major challenge for human neuroscience. Extensive work available from animal brain models unfortunately does not translate in a straightforward way to human subcortical anatomy nor does it shed much light on its involvement in human cognition (Steiner and Tseng, 2017). Besides serious difficulties in obtaining adequate measures of subcortical neural activity in functional MRI (de Hollander et al., 2017; Miletić et al., 2020), atlases and techniques for labeling accurately and reliably individual subcortical structures have also been scarce (Frazier et al., 2005; Chakravarty et al., 2006; Ahsan et al., 2007; Yelnik et al., 2007; Qiu et al., 2010; Patenaude et al., 2011), typically labeling the thalamus, striatum (or its subdivision into caudate and putamen), and globus pallidus (internal and external segments combined), sometimes the amygdala. However, recent advances in anatomical MRI, combining multiple contrasts and/or quantitative MRI mapping and utilizing the higher resolution achievable with 7 Tesla (7T) and above have started to reduce the gap, each mapping a few additional structures or sub-structures, primarily the iron-rich substantia nigra, red nucleus and sub-thalamic nucleus (Keuken et al., 2013; Xiao et al., 2015; Visser et al., 2016a; Visser et al., 2016b; Wang et al., 2016; Makowski et al., 2018; Ewert et al., 2018; Iglesias et al., 2018; Pauli et al., 2018; Sitek et al., 2019). While these efforts generated valuable atlases, they do not yet enable to identify many subcortical structures in individual subjects. Manual delineation, on the other hand, requires extensive manual labor from highly trained experts which cannot be easily applied to large cohorts or clinical settings.

Here, we propose a new automated parcellation technique to identify and label 17 individual subcortical structures of varying size and composition in individual subjects, based on a large quantitative 7T MRI database (Alkemade et al., 2020), using quantitative maps of relaxation rates R1 and R2* (1/T1 and 1/T2*, respectively) and quantitative susceptibility maps (QSM) as anatomical contrasts. The algorithm, named Multi-contrast Anatomical Subcortical Structure Parcellation (MASSP), follows a Bayesian multi-object approach similar in essence to previous efforts (Fischl et al., 2002; Eugenio Iglesias et al., 2013; Visser et al., 2016a; Garzón et al., 2018), combining shape priors, intensity distribution models, spatial relationships, and global constraints. The main innovation of our approach is to explicitly estimate interfaces between subcortical structures based on a joint model derived from signed distance functions. Modeling interfaces in addition to the structure itself provides a rich basis to encode relationships and anatomical knowledge in shape and intensity priors. A voxel-wise Markovian diffusion regularizes the combined priors for each defined interface, lowering the imaging noise. Finally, the voxel-wise posteriors for the different structures and interfaces are further combined into global anatomical parcels by topology correction and region growing taking into account volumetric priors, which regularizes parcellation results further in smaller nuclei with low or heterogeneous contrast. To validate the results from this new method, in a thorough comparison with expert manual labeling, we show that the proposed method provides results very close from manual raters in many structures and exhibit reasonable bias across the adult lifespan. The method can easily be extended to new structures, can be applied to any quantitative MRI dataset and is available in Open Source as part of Nighres (Huntenburg et al., 2018), a neuroimage analysis package aimed at high-resolution neuroimaging.

The MASSP parcellation method presented here has been trained to parcellate the following 17 structures: striatum (Str), thalamus (Tha), lateral, 3rd and 4th ventricles (LV, 3V, 4V), amygdala (Amg), globus pallidus internal segment (GPi) and external segment (GPe), SN, STN, red nucleus (RN), ventral tegmental area (VTA), fornix (fx), internal capsule (ic), periaqueductal gray (PAG), pedunculopontine nucleus (PPN), and claustrum (Cl), see Figure 1. These structures include the most commonly defined subcortical regions (Str, Tha, Amg, LV), the main iron-rich nuclei (GPi, GPe, RN, SN, STN), as well as smaller, less studied areas (VTA, PAG, PPN, Cl), white matter structures (ic, fx), and the central ventricles (3V, 4V).

Figure 1

Download asset Open asset

MASSP uses a data set of ten expert delineations as a basis for its modeling. From the delineations, an atlas of interfaces between structures, shape skeletons, and interface intensity histograms are generated, and used as prior in a multiple-step non-iterative Bayesian algorithm, see Figure 2 and Materials and methods.

Figure 2

Download asset Open asset

Validation against manual delineations

In a leave-one-out validation study comparing performance with the manual delineations, MASSP performed above 95% of the level of quality of the raters for Str, Tha, 4V, GPe, SN, RN, VTA, ic in terms of Dice overlap, the most stringent of the quality measures (see Figures 3 and 4 and Table 1). Several of the smaller structures have lower overlap ratios likely due to their smaller size (GPi, STN, PAG, PPN). Structures with an elongated shape (fx, Cl) remain challenging, due to the fact that small differences in location can substantially reduce overlap (Bazin et al., 2016). Despite these challenges, when comparing the dilated Dice scores, all structures were above 75% of overlap, with most reaching over 90% of the manual raters ability. Note that the Dice coefficient is very sensitive to size, as smaller structures will have lower overlap ratios for the same number of misclassified voxels. The dilated Dice coefficient is more representative of the variability regardless of size, as the smaller structures can reach high levels of overlap, both in manual and automated parcellations (see Table 1). The average surface distance confirms these results, showing values generally between one and two voxels of distance at a resolution of 0.7 mm, except in the cases of Amg, LV, fx, PPN, and Cl. These structures are generally more variable (LV), elongated (fx, Cl), or have a particularly low contrast with neighboring regions (Amg, PPN).

Table 1

	Str	STN	SN	RN	GPi	GPe	Tha	LV	3V	4V	Amg	ic	VTA	fx	PAG	PPN	Cl
Dice overlap
MASSP	0.893	0.648	0.805	0.870	0.702	0.800	0.867	0.849	0.741	0.869	0.723	0.745	0.570	0.527	0.641	0.496	0.485
Manual	0.897	0.800	0.841	0.875	0.762	0.813	0.877	0.907	0.797	0.882	0.866	0.732	0.574	0.823	0.791	0.665	0.727
Ratio	0.995	0.811	0.957	0.996	0.925	0.987	0.989	0.936	0.936	0.988	0.836	1.020	0.994	0.641	0.814	0.754	0.664
Dilated overlap
MASSP	0.982	0.919	0.977	0.991	0.909	0.956	0.970	0.929	0.890	0.951	0.891	0.915	0.863	0.756	0.897	0.795	0.789
Manual	0.987	0.988	0.985	0.995	0.953	0.972	0.970	0.967	0.944	0.961	0.978	0.924	0.818	0.957	0.960	0.910	0.914
Ratio	0.995	0.930	0.992	0.995	0.955	0.984	1.000	0.961	0.946	0.991	0.911	0.992	1.059	0.790	0.935	0.879	0.863
Average surface distance
MASSP	0.750	0.911	0.676	0.491	1.140	0.863	1.058	2.690	0.994	0.817	1.476	1.275	1.074	2.950	0.955	1.484	1.685
Manual	0.723	0.508	0.571	0.482	0.902	0.804	0.971	0.615	0.637	0.671	0.779	1.045	1.204	0.703	0.555	0.801	0.670
Ratio	0.971	0.590	0.852	0.996	0.861	0.943	0.916	0.277	0.662	1.020	0.553	0.834	1.161	0.287	0.610	0.619	0.465

Figure 3

Download asset Open asset

Figure 4

Download asset Open asset

Application to new MRI contrasts

Quantitative MRI has only become recently applicable in larger studies, thanks in part to the development of integrated multi-parameter sequences (Weiskopf et al., 2013; Caan et al., 2019). Many data sets, including large-scale open databases, use more common T1- and T2-weighted MRI. In order to test the applicability of MASSP to such contrasts, we obtained the test-retest subset of the Human Connectome Project (HCP, Van Essen et al., 2013) and applied MASSP to the 45 pre-processed and skull-stripped T1- and T2-weighted images from each of the two test and retest sessions. While performing manual delineations on the new contrasts would be preferable, the model is already rich enough to provide stable parcellations. Test-retest reproducibility is similarly high for MASSP and Freesurfer, and are generally in agreement, see Figure 5 and Table 3.

Table 3

	Str	STN	SN	RN	GPi	GPe	Tha	LV	3V	4V	Amg	ic	VTA	fx	PAG	PPN	Cl
Dice overlap
MASSP test-retest	0.914	0.701	0.818	0.829	0.791	0.859	0.928	0.881	0.837	0.870	0.866	0.860	0.738	0.774	0.714	0.713	0.785
Freesurfer test-retest	0.898				0.770		0.919	0.894	0.842	0.849	0.852
MASSP – Freesurfer	0.876				0.778		0.838	0.858	0.430	0.769	0.692
Dilated overlap
MASSP test-retest	0.987	0.939	0.977	0.978	0.963	0.977	0.990	0.980	0.979	0.986	0.981	0.973	0.969	0.961	0.965	0.972	0.966
Freesurfer test-retest	0.986				0.926		0.986	0.989	0.972	0.975	0.978
MASSP – Freesurfer	0.954				0.788		0.919	0.934	0.435	0.901	0.866
Average surface distance
MASSP test-retest	0.513	0.528	0.467	0.461	0.532	0.508	0.488	0.509	0.391	0.419	0.533	0.536	0.431	0.464	0.428	0.402	0.433
Freesurfer test-retest	0.876				0.778		0.838	0.858	0.430	0.769	0.692
MASSP – Freesurfer	0.976				1.673		1.605	1.946	5.699	1.428	1.478

Figure 5

Download asset Open asset

Biases due to atlas size

A common concern of brain parcellation methods is the risk of biases, as they are typically built from a small number of manual delineations. Our data set is part of a large scale study of the subcortex, for which we obtained manual delineations of the STN, SN, RN, GPe, and GPi on 105 subjects over the adult lifespan (18–80 year old, see Alkemade et al., 2020 for details). First, we investigated the impact of atlas size. We randomly assigned half of the subjects from each decade to two groups, and built atlas priors from subsets of 3, 5, 8, 10, 12, 15, and 18 subjects from the first group. The subjects used in the atlas were taken randomly from each decade (18-30, 31-40, 41-50,51-60, 61-70, 71-80), so as to maximize the age range represented in each atlas. Atlases of increasing size were constructed by adding subjects to previous atlases, so that atlases of increasing complexity include all subjects from simpler atlases. Results applying these atlases to parcellate the second group are given in Figure 6. As in previous studies (Eugenio Iglesias et al., 2013; Bazin and Pham, 2008), performance quickly stabilized with atlases of more than five subjects (no significant difference in Welch’s t-tests between using 18 subjects or any subset of 8 or more for all structures and measures).

Figure 6

Download asset Open asset

Biases due to age differences

To more specifically test the influence of age on parcellation accuracy, we defined again six age groups by decade and randomly selected 10 subjects from each group. Each set of subjects was used as priors for the five structures above, and applied to the other age groups. Results are summarized in Figure 7. Examining this age bias, we can see a decrease in performance when parcellating subjects in the range of 60 to 80 years of age. The choice of priors seem to have a limited impact, which varies across structures. In particular, using priors from a similar age group is not always beneficial.

Figure 7

Download asset Open asset

Bias on individual measures

Finally, we investigated the impact of this decrease in performance in the estimation of anatomical quantities, see Figure 8. The bias did affect the morphometric measures of structure volume and thickness, but the effects on the local measure of thickness was reduced compared to the global measure of volume. Quantitative MRI averages were very stable even when age biases are present in the parcellations.

Figure 8

Download asset Open asset

For reference, we report structure volumes, thickness, R1, R2* and QSM values estimated from the entire AHEAD cohort for different age groups, extending our previous work based on manual delineations on a different data set (Keuken et al., 2017; Forstmann et al., 2014). Results are given in Table 4, describing average volumes, thickness, and quantitative MRI parameters for young, middle-aged, and older subjects for the 17 subcortical structures.

Table 4

Age	Str	STN	SN	RN	GPi	GPe	Tha	LV	3V	4V	Amg	ic	VTA	fx	PAG	PPN	Cl
Volume ($m{m}^3$)
18-40	10656	118	566	253	567	1366	7112	7524	1895	1391	1315	4335	254	1632	250	193	843
41-60	10572	124	583	256	586	1403	7492	8850	2024	1408	1363	4495	264	1808	255	195	830
61-80	10734	130	584	260	586	1397	7463	9142	2023	1407	1321	4407	272	1910	259	192	829
Thickness ($mm$)
18-40	5.94	1.89	2.55	4.64	3.09	3.56	8.31	4.27	2.77	4.03	4.81	4.06	1.69	2.06	1.78	1.92	1.79
41-60	5.47	1.86	2.66	4.58	2.96	3.41	8.28	5.08	2.95	3.89	4.85	4.19	1.76	1.96	1.84	1.86	1.80
61-80	5.22	1.83	2.60	4.11	2.92	3.22	8.28	4.90	3.18	4.06	4.73	4.19	1.80	1.97	1.90	1.95	1.82
qR1 ($Hz$)
18-40	0.647	0.949	0.857	0.928	0.868	0.850	0.761	0.332	0.346	0.274	0.546	0.906	0.819	0.714	0.654	0.779	0.650
41-60	0.662	0.968	0.893	0.939	0.879	0.856	0.758	0.278	0.315	0.269	0.559	0.904	0.833	0.671	0.653	0.771	0.664
61-80	0.648	0.952	0.882	0.903	0.860	0.830	0.743	0.273	0.300	0.270	0.552	0.884	0.814	0.638	0.647	0.764	0.669
qR2* ($Hz$)
18-40	43.8	67.1	67.8	63.2	75.2	79.6	38.1	14.7	18.9	9.0	25.5	36.8	39.2	37.4	25.9	32.7	32.6
41-60	50.4	74.1	74.1	77.1	80.2	87.9	40.3	8.4	12.4	11.7	28.1	38.7	42.8	37.4	28.0	33.4	36.9
61-80	51.8	77.0	72.5	73.8	77.8	87.0	40.1	8.5	10.2	12.0	30.1	39.6	52.6	35.7	28.4	34.2	35.4
QSM ($ppm$)
18-40	0.0329	0.0609	0.0738	0.0717	0.1015	0.1150	0.0138	0.0130	0.0100	0.0279	0.0036	−0.0234	0.0241	0.0079	0.0119	0.0135	−0.0122
41-60	0.0400	0.0647	0.0713	0.0829	0.0984	0.1241	0.0134	0.0115	0.0025	0.0234	0.0085	−0.0226	0.0201	0.0079	0.0089	0.0099	−0.0110
61-80	0.0411	0.0705	0.0610	0.0738	0.0925	0.1249	0.0064	0.0089	−0.0034	0.0236	0.0061	−0.0243	0.0177	0.0100	0.0039	0.0096	−0.0091

Our goal with the MASSP algorithm was to provide a fully automated method to delineate as many subcortical structures as possible on high-resolution structural MRI now available on 7T scanners. We modeled 17 distinct structures, taking into account location, shape, volume, and quantitative MRI contrasts to provide individual subject parcellations. Based on our results, we can be confident that the automated parcellation technique performs comparably to human experts, providing delineations within one or two voxels of the structure boundaries (dilated Dice overlap over 75% for all structures, including in aging groups). Results were nearly indistinguishable from expert delineations for eight major structures (Str, Tha, 4V, GPe, SN, RN, VTA, ic), and smaller structures retain high levels of overlap, comparable to trained human raters. This parcellation includes the most commonly defined structures (Str, Tha, SN, RN, STN) with overlap scores comparable to those previously reported (Garzón et al., 2018; Visser et al., 2016a; Eugenio Iglesias et al., 2013; Chakravarty et al., 2013; Patenaude et al., 2011). More importantly, it also includes structures seldom or never before considered in MRI atlases and parcellation methods, such as GPe, GPi, VTA, 3V, 4V, ic, fx, PAG, PPN, Cl. The technique handles structures of varying sizes well, as indicated by dilated overlap and boundary distance. Additional structures can be added, if they can be reliably delineated by expert raters on single-subject MRI at achievable resolutions. Some enhancement techniques such as building a multi-subject template (Pauli et al., 2018) or adding a denoising step (Bazin et al., 2019) may be beneficial. Co-registration to a high-precision atlas as in Ewert et al., 2018 may also improve the initial alignment over the MASSP group average template.

Age biases are present both in expert manual delineations and automated parcellation techniques. Age trajectories in volume and quantitative MR parameters indicate systematic shifts in contrast intensities and an increasing variability with age, associated with changing myelination, iron deposition, and brain atrophy (Draganski et al., 2011; Daugherty and Raz, 2013; Fjell et al., 2013; Keuken et al., 2017). These changes seem only to impact the parcellation accuracy for age groups beyond 60 years and age-matched priors did not provide specific improvements, thus indicating that an explicit modeling of age effects may be required to further improve parcellation quality in elderly populations. These results also point to exercising caution when applying automated parcellation methods to study morphometry in elderly or diseased populations, where measured differences may include biases. They also point out that while global volume and local thickness are indeed affected by such biases, quantitative MRI measures are much more robust. Note that this bias is likely present is many automated methods, although they have not been systematically investigated due to the extensive manual labor required. Interestingly, biases also exist in expert delineations: when the size or shape of a structure is refined in neuroanatomical studies, experts may become more or less conservative in their delineations. Automated methods provide a more objective measure in such case, as the source of their bias is explicitly encoded in the atlas prior delineations and computational model. Important applications of subcortical parcellation also include deep-brain stimulation surgery (Ewert et al., 2018), where the number of structures parcellated by MASSP can help neurosurgeons orient themselves more easily, although precise targeting will still require manual refinements, especially in neurodegenerative diseases.

We observed that dilated overlap, that is, the overlap of structures up to one voxel, provided a measure of accuracy largely independent of size, for automated or manual delineations. Imprecision in the range of one voxel in the boundary is to be expected due to partial voluming which impacts Dice overlap. The dilated overlap measure is a better representative of performance and indicates that conservative or inclusive versions of the subcortical regions can be obtained by eroding or dilating the estimated boundary by a single voxel. Such masks may be useful when separating functional MRI signals between neighboring nuclei or when locating smaller features inside a structure. Additionally, the Bayesian estimation framework provides voxel-wise probability values, which can also be used to further weight the contribution of each voxel within a region in subsequent analyses.

In summary, our method provides fast and accurate parcellation for subcortical structures of varying size, taking advantage of the high resolution offered by 7T and the specificity of quantitative MRI. The algorithm is based on an explicit model of structures given in a Bayesian framework and is free of tuning parameters. Given a different set of regions of interest or different populations, new priors can be automatically generated and used as the basis for the algorithm. If more MRI contrasts are available, the method can also be augmented to take them into account. The main requirement for the technique is a set of manual delineations of all the structures of interest in a small group of representative subjects. Performance may further improve with the number of included structures, as the number of distinct interfaces increases, refining in particular the intensity priors. In future works, we plan to include more structures or sub-structures and model the effects of age on the priors. We hope that the method, available in open source, will help neuroscience researchers to include more subcortical regions in their structural and functional imaging studies.

The tool presented in this article is available in open source on Github (https://github.com/nighres/nighres). The atlases necessary to run the algorithm have been deposited on the University of Amsterdam FigShare (https://doi.org/10.21942/uva.12074175.v1 and https://doi.org/10.21942/uva.12301106.v2). A single sample subject data set has been deposited on the University of Amsterdam FigShare (https://doi.org/10.21942/uva.12280316.v2). All the measurements used to generate the figures included in the article have been deposited on the University of Amsterdam FigShare (https://doi.org/10.21942/uva.12452444.v1).

1. 1. Alkemade A
   2. Mulder MJ
   3. Groot JM
   4. Isaacs BR
   5. van Berendonk N
   6. Lute N
   7. Isherwood SJ
   8. Bazin P-L
   9. Forstmann BU

   (2020) FigShare / University of Amsterdam / Amsterdam University of Applied Sciences

   The Amsterdam Ultra-high field adult lifespan database (AHEAD): A freely available multimodal 7 Tesla submillimeter magnetic resonance imaging database.

   https://doi.org/10.21942/uva.10007840.v1
2. 1. Van Essen DC
   2. Smith SM
   3. Barch DM
   4. Behrens TEJ
   5. Yacoub E
   6. Ugurbil K

   (2013) WU-Minn HCP Retest Data

   ID hcp-young-adult. The WU-Minn Human Connectome Project.

   https://www.humanconnectome.org/study/hcp-young-adult

1. 1. Ahsan RL
   2. Allom R
   3. Gousias IS
   4. Habib H
   5. Turkheimer FE
   6. Free S
   7. Lemieux L
   8. Myers R
   9. Duncan JS
   10. Brooks DJ
   11. Koepp MJ
   12. Hammers A

   (2007) Volumes, spatial extents and a probabilistic atlas of the human basal ganglia and thalamus

   NeuroImage 38:261–270.

   https://doi.org/10.1016/j.neuroimage.2007.06.004

   • PubMed
   • Google Scholar
2. 1. Alkemade A
   2. Mulder MJ
   3. Groot JM
   4. Isaacs BR
   5. van Berendonk N
   6. Lute N
   7. Isherwood SJ
   8. Bazin PL
   9. Forstmann BU

   (2020) The Amsterdam Ultra-high field adult lifespan database (AHEAD): A freely available multimodal 7 tesla submillimeter magnetic resonance imaging database

   NeuroImage 221:117200.

   https://doi.org/10.1016/j.neuroimage.2020.117200

   • PubMed
   • Google Scholar
3. 1. Andersen HH
   2. Johnsen KB
   3. Moos T

   (2014) Iron deposits in the chronically inflamed central nervous system and contributes to neurodegeneration

   Cellular and Molecular Life Sciences 71:1607–1622.

   https://doi.org/10.1007/s00018-013-1509-8

   • PubMed
   • Google Scholar
4. 1. Avants BB
   2. Epstein CL
   3. Grossman M
   4. Gee JC

   (2008) Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain

   Medical Image Analysis 12:26–41.

   https://doi.org/10.1016/j.media.2007.06.004

   • PubMed
   • Google Scholar
5. Book

   1. Bazin P-L
   2. Ellingsen LM
   3. Pham DL

   (2007) Digital Homeomorphisms in Deformable Registration

   In: Karssemeijer N, Lelieveldt B, editors. Information Processing in Medical Imaging. Berlin, Heidelberg: Springer Berlin Heidelberg. pp. 211–222.

   https://doi.org/10.1007/978-3-540-73273-0_18

   • Google Scholar
6. Conference

   1. Bazin P-L
   2. Plessis V
   3. Fan AP
   4. Villringer A
   5. Gauthier CJ

   (2016) Vessel segmentation from quantitative susceptibility maps for local oxygenation venography

   2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI). pp. 1135–1138.

   https://doi.org/10.1109/ISBI.2016.7493466

   • Google Scholar
7. 1. Bazin PL
   2. Alkemade A
   3. van der Zwaag W
   4. Caan M
   5. Mulder M
   6. Forstmann BU

   (2019) Denoising High-Field Multi-Dimensional MRI with local complex PCA

   Frontiers in Neuroscience 13:66.

   https://doi.org/10.3389/fnins.2019.01066

   • PubMed
   • Google Scholar
8. 1. Bazin PL
   2. Pham DL

   (2007) Topology correction of segmented medical images using a fast marching algorithm

   Computer Methods and Programs in Biomedicine 88:182–190.

   https://doi.org/10.1016/j.cmpb.2007.08.006

   • PubMed
   • Google Scholar
9. 1. Bazin PL
   2. Pham DL

   (2008) Homeomorphic brain image segmentation with topological and statistical atlases

   Medical Image Analysis 12:616–625.

   https://doi.org/10.1016/j.media.2008.06.008

   • PubMed
   • Google Scholar
10. 1. Blum H

    (1973) Biological shape and visual science. I

    Journal of Theoretical Biology 38:205–287.

    https://doi.org/10.1016/0022-5193(73)90175-6

    • PubMed
    • Google Scholar
11. 1. Caan MWA
    2. Bazin P
    3. Marques JP
    4. Hollander G
    5. Dumoulin SO
    6. Zwaag W

    (2019) MP2RAGEME: T 1 , T 2 * , and QSM mapping in one sequence at 7 tesla

    Human Brain Mapping 40:1786–1798.

    https://doi.org/10.1002/hbm.24490

    • PubMed
    • Google Scholar
12. 1. Chakravarty MM
    2. Bertrand G
    3. Hodge CP
    4. Sadikot AF
    5. Collins DL

    (2006) The creation of a brain atlas for image guided neurosurgery using serial histological data

    NeuroImage 30:359–376.

    https://doi.org/10.1016/j.neuroimage.2005.09.041

    • PubMed
    • Google Scholar
13. 1. Chakravarty MM
    2. Steadman P
    3. van Eede MC
    4. Calcott RD
    5. Gu V
    6. Shaw P
    7. Raznahan A
    8. Collins DL
    9. Lerch JP

    (2013) Performing label-fusion-based segmentation using multiple automatically generated templates: maget brain: label fusion segmentation using automatically generated templates

    Human Brain Mapping 34:2635–2654.

    https://doi.org/10.1002/hbm.22092

    • PubMed
    • Google Scholar
14. 1. Daugherty A
    2. Raz N

    (2013) Age-related differences in iron content of subcortical nuclei observed in vivo: a meta-analysis

    NeuroImage 70:113–121.

    https://doi.org/10.1016/j.neuroimage.2012.12.040

    • PubMed
    • Google Scholar
15. 1. de Hollander G
    2. Keuken MC
    3. van der Zwaag W
    4. Forstmann BU
    5. Trampel R

    (2017) Comparing functional MRI protocols for small, iron-rich basal ganglia nuclei such as the subthalamic nucleus at 7 T and 3 T

    Human Brain Mapping 38:3226–3248.

    https://doi.org/10.1002/hbm.23586

    • PubMed
    • Google Scholar
16. 1. Dice LR

    (1945) Measures of the amount of ecologic association between species

    Ecology 26:297–302.

    https://doi.org/10.2307/1932409

    • Google Scholar
17. 1. Draganski B
    2. Ashburner J
    3. Hutton C
    4. Kherif F
    5. Frackowiak RS
    6. Helms G
    7. Weiskopf N

    (2011) Regional specificity of MRI contrast parameter changes in normal ageing revealed by voxel-based quantification (VBQ)

    NeuroImage 55:1423–1434.

    https://doi.org/10.1016/j.neuroimage.2011.01.052

    • PubMed
    • Google Scholar
18. 1. Eugenio Iglesias J
    2. Rory Sabuncu M
    3. Van Leemput K

    (2013) A unified framework for cross-modality multi-atlas segmentation of brain MRI

    Medical Image Analysis 17:1181–1191.

    https://doi.org/10.1016/j.media.2013.08.001

    • PubMed
    • Google Scholar
19. 1. Ewert S
    2. Plettig P
    3. Li N
    4. Chakravarty MM
    5. Collins DL
    6. Herrington TM
    7. Kühn AA
    8. Horn A

    (2018) Toward defining deep brain stimulation targets in MNI space: a subcortical atlas based on multimodal MRI, histology and structural connectivity

    NeuroImage 170:271–282.

    https://doi.org/10.1016/j.neuroimage.2017.05.015

    • PubMed
    • Google Scholar
20. 1. Ewert S
    2. Horn A
    3. Finkel F
    4. Li N
    5. Kühn AA
    6. Herrington TM

    (2019) Optimization and comparative evaluation of nonlinear deformation algorithms for atlas-based segmentation of DBS target nuclei

    NeuroImage 184:586–598.

    https://doi.org/10.1016/j.neuroimage.2018.09.061

    • PubMed
    • Google Scholar
21. 1. Fischl B
    2. Salat DH
    3. Busa E
    4. Albert M
    5. Dieterich M
    6. Haselgrove C
    7. van der Kouwe A
    8. Killiany R
    9. Kennedy D
    10. Klaveness S
    11. Montillo A
    12. Makris N
    13. Rosen B
    14. Dale AM

    (2002) Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain

    Neuron 33:341–355.

    https://doi.org/10.1016/s0896-6273(02)00569-x

    • PubMed
    • Google Scholar
22. 1. Fjell AM
    2. Westlye LT
    3. Grydeland H
    4. Amlien I
    5. Espeseth T
    6. Reinvang I
    7. Raz N
    8. Holland D
    9. Dale AM
    10. Walhovd KB
    11. Alzheimer Disease Neuroimaging Initiative

    (2013) Critical ages in the life course of the adult brain: nonlinear subcortical aging

    Neurobiology of Aging 34:2239–2247.

    https://doi.org/10.1016/j.neurobiolaging.2013.04.006

    • PubMed
    • Google Scholar
23. 1. Fonov V
    2. Evans AC
    3. Botteron K
    4. Almli CR
    5. McKinstry RC
    6. Collins DL
    7. Brain Development Cooperative Group

    (2011) Unbiased average age-appropriate atlases for pediatric studies

    NeuroImage 54:313–327.

    https://doi.org/10.1016/j.neuroimage.2010.07.033

    • PubMed
    • Google Scholar
24. 1. Forstmann BU
    2. Keuken MC
    3. Schafer A
    4. Bazin PL
    5. Alkemade A
    6. Turner R

    (2014) Multi-modal ultra-high resolution structural 7-Tesla MRI data repository

    Scientific Data 1:140050.

    https://doi.org/10.1038/sdata.2014.50

    • PubMed
    • Google Scholar
25. 1. Forstmann BU
    2. de Hollander G
    3. van Maanen L
    4. Alkemade A
    5. Keuken MC

    (2016) Towards a mechanistic understanding of the human subcortex

    Nature Reviews Neuroscience 18:57–65.

    https://doi.org/10.1038/nrn.2016.163

    • PubMed
    • Google Scholar
26. 1. Forstmann BU
    2. Isaacs BR
    3. Temel Y

    (2017) Ultra high field MRI-Guided deep brain stimulation

    Trends in Biotechnology 35:904–907.

    https://doi.org/10.1016/j.tibtech.2017.06.010

    • PubMed
    • Google Scholar
27. 1. Frazier JA
    2. Chiu S
    3. Breeze JL
    4. Makris N
    5. Lange N
    6. Kennedy DN
    7. Herbert MR
    8. Bent EK
    9. Koneru VK
    10. Dieterich ME
    11. Hodge SM
    12. Rauch SL
    13. Grant PE
    14. Cohen BM
    15. Seidman LJ
    16. Caviness VS
    17. Biederman J

    (2005) Structural brain magnetic resonance imaging of limbic and thalamic volumes in pediatric bipolar disorder

    American Journal of Psychiatry 162:1256–1265.

    https://doi.org/10.1176/appi.ajp.162.7.1256

    • PubMed
    • Google Scholar
28. 1. Garzón B
    2. Sitnikov R
    3. Bäckman L
    4. Kalpouzos G

    (2018) Automated segmentation of midbrain structures with high iron content

    NeuroImage 170:199–209.

    https://doi.org/10.1016/j.neuroimage.2017.06.016

    • PubMed
    • Google Scholar
29. 1. Herting MM
    2. Johnson C
    3. Mills KL
    4. Vijayakumar N
    5. Dennison M
    6. Liu C
    7. Goddings AL
    8. Dahl RE
    9. Sowell ER
    10. Whittle S
    11. Allen NB
    12. Tamnes CK

    (2018) Development of subcortical volumes across adolescence in males and females: a multisample study of longitudinal changes

    NeuroImage 172:194–205.

    https://doi.org/10.1016/j.neuroimage.2018.01.020

    • PubMed
    • Google Scholar
30. 1. Hill RA
    2. Li AM
    3. Grutzendler J

    (2018) Lifelong cortical myelin plasticity and age-related degeneration in the live mammalian brain

    Nature Neuroscience 21:683–695.

    https://doi.org/10.1038/s41593-018-0120-6

    • PubMed
    • Google Scholar
31. 1. Ho TC
    2. Gutman B
    3. Pozzi E
    4. Grabe HJ
    5. Hosten N
    6. Wittfeld K
    7. Völzke H
    8. Baune B
    9. Dannlowski U
    10. Förster K
    11. Grotegerd D
    12. Redlich R
    13. Jansen A
    14. Kircher T
    15. Krug A
    16. Meinert S
    17. Nenadic I
    18. Opel N
    19. Dinga R
    20. Veltman DJ
    21. Schnell K
    22. Veer I
    23. Walter H
    24. Gotlib IH
    25. Sacchet MD
    26. Aleman A
    27. Groenewold NA
    28. Stein DJ
    29. Li M
    30. Walter M
    31. Ching CRK
    32. Jahanshad N
    33. Ragothaman A
    34. Isaev D
    35. Zavaliangos‐Petropulu A
    36. Thompson PM
    37. Sämann PG
    38. Schmaal L

    (2020) Subcortical shape alterations in major depressive disorder: findings from the ENIGMA major depressive disorder working group

    Human Brain Mapping 1:e24988.

    https://doi.org/10.1002/hbm.24988

    • Google Scholar
32. 1. Huntenburg JM
    2. Steele CJ
    3. Bazin PL

    (2018) Nighres: processing tools for high-resolution neuroimaging

    GigaScience 7:giy082.

    https://doi.org/10.1093/gigascience/giy082

    • PubMed
    • Google Scholar
33. 1. Iglesias JE
    2. Insausti R
    3. Lerma-Usabiaga G
    4. Bocchetta M
    5. Van Leemput K
    6. Greve DN
    7. van der Kouwe A
    8. Fischl B
    9. Caballero-Gaudes C
    10. Paz-Alonso PM
    11. Alzheimer's Disease Neuroimaging Initiative

    (2018) A probabilistic atlas of the human thalamic nuclei combining ex vivo MRI and histology

    NeuroImage 183:314–326.

    https://doi.org/10.1016/j.neuroimage.2018.08.012

    • PubMed
    • Google Scholar
34. 1. Ji JL
    2. Spronk M
    3. Kulkarni K
    4. Repovš G
    5. Anticevic A
    6. Cole MW

    (2019) Mapping the human brain's cortical-subcortical functional network organization

    NeuroImage 185:35–57.

    https://doi.org/10.1016/j.neuroimage.2018.10.006

    • PubMed
    • Google Scholar
35. 1. Johansen-Berg H

    (2013) Human connectomics - what will the future demand?

    NeuroImage 80:541–544.

    https://doi.org/10.1016/j.neuroimage.2013.05.082

    • PubMed
    • Google Scholar
36. 1. Keuken MC
    2. Bazin PL
    3. Schäfer A
    4. Neumann J
    5. Turner R
    6. Forstmann BU

    (2013) Ultra-high 7T MRI of structural age-related changes of the subthalamic nucleus

    Journal of Neuroscience 33:4896–4900.

    https://doi.org/10.1523/JNEUROSCI.3241-12.2013

    • PubMed
    • Google Scholar
37. 1. Keuken MC
    2. Bazin PL
    3. Backhouse K
    4. Beekhuizen S
    5. Himmer L
    6. Kandola A
    7. Lafeber JJ
    8. Prochazkova L
    9. Trutti A
    10. Schäfer A
    11. Turner R
    12. Forstmann BU

    (2017) Effects of aging on T₁, T₂*, and QSM MRI values in the subcortex

    Brain Structure & Function 222:2487–2505.

    https://doi.org/10.1007/s00429-016-1352-4

    • PubMed
    • Google Scholar
38. 1. Koshiyama D
    2. Fukunaga M
    3. Okada N
    4. Yamashita F
    5. Yamamori H
    6. Yasuda Y
    7. Fujimoto M
    8. Ohi K
    9. Fujino H
    10. Watanabe Y
    11. Kasai K
    12. Hashimoto R

    (2018) Role of subcortical structures on cognitive and social function in schizophrenia

    Scientific Reports 8:1183.

    https://doi.org/10.1038/s41598-017-18950-2

    • PubMed
    • Google Scholar
39. 1. Li W
    2. Avram AV
    3. Wu B
    4. Xiao X
    5. Liu C

    (2014) Integrated Laplacian-based phase unwrapping and background phase removal for quantitative susceptibility mapping

    NMR in Biomedicine 27:219–227.

    https://doi.org/10.1002/nbm.3056

    • PubMed
    • Google Scholar
40. 1. Makowski C
    2. Béland S
    3. Kostopoulos P
    4. Bhagwat N
    5. Devenyi GA
    6. Malla AK
    7. Joober R
    8. Lepage M
    9. Chakravarty MM

    (2018) Evaluating accuracy of striatal, pallidal, and thalamic segmentation methods: comparing automated approaches to manual delineation

    NeuroImage 170:182–198.

    https://doi.org/10.1016/j.neuroimage.2017.02.069

    • PubMed
    • Google Scholar
41. 1. Maquer G
    2. Musy SN
    3. Wandel J
    4. Gross T
    5. Zysset PK

    (2015) Bone volume fraction and fabric anisotropy are better determinants of trabecular bone stiffness than other morphological variables

    Journal of Bone and Mineral Research 30:1000–1008.

    https://doi.org/10.1002/jbmr.2437

    • PubMed
    • Google Scholar
42. 1. Marquand AF
    2. Haak KV
    3. Beckmann CF

    (2017) Functional corticostriatal connection topographies predict goal directed behaviour in humans

    Nature Human Behaviour 1:0146.

    https://doi.org/10.1038/s41562-017-0146

    • PubMed
    • Google Scholar
43. 1. Marques JP
    2. Kober T
    3. Krueger G
    4. van der Zwaag W
    5. Van de Moortele PF
    6. Gruetter R

    (2010) MP2RAGE, a self bias-field corrected sequence for improved segmentation and T1-mapping at high field

    NeuroImage 49:1271–1281.

    https://doi.org/10.1016/j.neuroimage.2009.10.002

    • PubMed
    • Google Scholar
44. 1. McCoy RC
    2. Wakefield J
    3. Akey JM

    (2017) Impacts of Neanderthal-Introgressed sequences on the landscape of human gene expression

    Cell 168:916–927.

    https://doi.org/10.1016/j.cell.2017.01.038

    • PubMed
    • Google Scholar
45. 1. Miletić S
    2. Bazin PL
    3. Weiskopf N
    4. van der Zwaag W
    5. Forstmann BU
    6. Trampel R

    (2020) fMRI protocol optimization for simultaneously studying small subcortical and cortical Areas at 7T

    NeuroImage 219:116992.

    https://doi.org/10.1016/j.neuroimage.2020.116992

    • PubMed
    • Google Scholar
46. 1. Mosley PE
    2. Smith D
    3. Coyne T
    4. Silburn P
    5. Breakspear M
    6. Perry A

    (2018) The site of stimulation moderates neuropsychiatric symptoms after subthalamic deep brain stimulation for Parkinson's disease

    NeuroImage: Clinical 18:996–1006.

    https://doi.org/10.1016/j.nicl.2018.03.009

    • PubMed
    • Google Scholar
47. 1. Patenaude B
    2. Smith SM
    3. Kennedy DN
    4. Jenkinson M

    (2011) A Bayesian model of shape and appearance for subcortical brain segmentation

    NeuroImage 56:907–922.

    https://doi.org/10.1016/j.neuroimage.2011.02.046

    • PubMed
    • Google Scholar
48. 1. Pauli WM
    2. Nili AN
    3. Tyszka JM

    (2018) A high-resolution probabilistic in vivo atlas of human subcortical brain nuclei

    Scientific Data 5:180063.

    https://doi.org/10.1038/sdata.2018.63

    • PubMed
    • Google Scholar
49. 1. Qiu A
    2. Brown T
    3. Fischl B
    4. Ma J
    5. Miller MI

    (2010) Atlas generation for subcortical and ventricular structures with its applications in shape analysis

    IEEE Transactions on Image Processing 19:1539–1547.

    https://doi.org/10.1109/TIP.2010.2042099

    • PubMed
    • Google Scholar
50. Preprint

    1. Sitek KR
    2. Gulban OF
    3. Calabrese E
    4. Johnson GA
    5. Lage-Castellanos A
    6. Moerel M
    7. Ghosh SS
    8. De Martino F

    (2019) Mapping the human subcortical auditory system using histology. post mortem MRI and in vivo MRI at 7T

    bioRxiv.

    https://doi.org/10.1101/568139

    • Google Scholar
51. 1. Smith SM

    (2002) Fast robust automated brain extraction

    Human Brain Mapping 17:143–155.

    https://doi.org/10.1002/hbm.10062

    • PubMed
    • Google Scholar
52. Book

    1. Steiner H
    2. Tseng KY

    (2017)

    Handbook of Basal Ganglia Structure and Function

    In: Steiner H, Tseng K, editors. Handbook of Behavioral Neuroscience, 24. Elsevier. pp. 1–1036.

    • Google Scholar
53. 1. Tardif CL
    2. Gauthier CJ
    3. Steele CJ
    4. Bazin PL
    5. Schäfer A
    6. Schaefer A
    7. Turner R
    8. Villringer A

    (2016) Advanced MRI techniques to improve our understanding of experience-induced neuroplasticity

    NeuroImage 131:55–72.

    https://doi.org/10.1016/j.neuroimage.2015.08.047

    • PubMed
    • Google Scholar
54. 1. Turner R

    (2019) Myelin and modeling: bootstrapping cortical microcircuits

    Frontiers in Neural Circuits 13:34.

    https://doi.org/10.3389/fncir.2019.00034

    • PubMed
    • Google Scholar
55. 1. Van Essen DC
    2. Smith SM
    3. Barch DM
    4. Behrens TE
    5. Yacoub E
    6. Ugurbil K
    7. WU-Minn HCP Consortium

    (2013) The WU-Minn human connectome project: an overview

    NeuroImage 80:62–79.

    https://doi.org/10.1016/j.neuroimage.2013.05.041

    • PubMed
    • Google Scholar
56. 1. Visser E
    2. Keuken MC
    3. Douaud G
    4. Gaura V
    5. Bachoud-Levi AC
    6. Remy P
    7. Forstmann BU
    8. Jenkinson M

    (2016a) Automatic segmentation of the striatum and globus pallidus using MIST: multimodal image segmentation tool

    NeuroImage 125:479–497.

    https://doi.org/10.1016/j.neuroimage.2015.10.013

    • PubMed
    • Google Scholar
57. 1. Visser E
    2. Keuken MC
    3. Forstmann BU
    4. Jenkinson M

    (2016b) Automated segmentation of the Substantia Nigra, subthalamic nucleus and red nucleus in 7T data at young and old age

    NeuroImage 139:324–336.

    https://doi.org/10.1016/j.neuroimage.2016.06.039

    • PubMed
    • Google Scholar
58. Conference

    1. Wang BT
    2. Poirier S
    3. Guo T
    4. Parrent AG
    5. Peters TM
    6. Khan AR
    7. Styner MA
    8. Angelini ED

    (2016)

    Generation and evaluation of an ultra-high-field atlas with applications in DBS planning

    Proceedings of SPIE Medical Imaging.

    • Google Scholar
59. 1. Weiskopf N
    2. Suckling J
    3. Williams G
    4. Correia MM
    5. Inkster B
    6. Tait R
    7. Ooi C
    8. Bullmore ET
    9. Lutti A

    (2013) Quantitative multi-parameter mapping of R1, PD(*), MT, and R2(*) at 3T: a multi-center validation

    Frontiers in Neuroscience 7:95.

    https://doi.org/10.3389/fnins.2013.00095

    • PubMed
    • Google Scholar
60. 1. Xiao Y
    2. Fonov V
    3. Bériault S
    4. Al Subaie F
    5. Chakravarty MM
    6. Sadikot AF
    7. Pike GB
    8. Collins DL

    (2015) Multi-contrast unbiased MRI atlas of a parkinson's disease population

    International Journal of Computer Assisted Radiology and Surgery 10:329–341.

    https://doi.org/10.1007/s11548-014-1068-y

    • PubMed
    • Google Scholar
61. 1. Yeatman JD
    2. Wandell BA
    3. Mezer AA

    (2014) Lifespan maturation and degeneration of human brain white matter

    Nature Communications 5:4932.

    https://doi.org/10.1038/ncomms5932

    • PubMed
    • Google Scholar
62. 1. Yelnik J
    2. Bardinet E
    3. Dormont D
    4. Malandain G
    5. Ourselin S
    6. Tandé D
    7. Karachi C
    8. Ayache N
    9. Cornu P
    10. Agid Y

    (2007) A three-dimensional, histological and deformable atlas of the human basal ganglia. I. Atlas construction based on immunohistochemical and MRI data

    NeuroImage 34:618–638.

    https://doi.org/10.1016/j.neuroimage.2006.09.026

    • PubMed
    • Google Scholar
63. 1. Zhang Y
    2. Wei H
    3. Cronin MJ
    4. He N
    5. Yan F
    6. Liu C

    (2018) Longitudinal atlas for normative human brain development and aging over the lifespan using quantitative susceptibility mapping

    NeuroImage 171:176–189.

    https://doi.org/10.1016/j.neuroimage.2018.01.008

    • PubMed
    • Google Scholar

Acceptance summary:

There was unanimous agreement that this method has strong potential to be a new "workhorse" tool in human neuroimaging that could substantially advance our ability to measure brain structures that are largely overlooked due to problems with segmentation.

Decision letter after peer review:

Thank you for submitting your article "Multi-contrast Anatomical Subcortical Structures Parcellation" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Timothy Verstynen as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Michael Frank as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Wolf-Julian Neumann (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). Specifically, we are asking editors to accept without delay manuscripts, like yours, that they judge can stand as eLife papers without additional data, even if they feel that they would make the manuscript stronger. Thus the revisions requested below only address clarity and presentation.

Summary:

In this study, Bazin and colleagues propose a novel segmentation algorithm for parcelling subcortical regions of the human brain that was developed from multiple MRI measures derived from the M2RAGEME sequence acquired on a 7T MRI system. The key advancement of this approach is a reliable segmentation of more subcortical areas (17 regions) in native space than what is possible with currently available methods. The authors validate their algorithm by comparing against age-related measures.

This manuscript was reviewed by three experts in the field, who found that this method has strong potential to be a new "workhorse" tool in human neuroimaging that could substantially advance our ability to measure brain structures that are largely overlooked due to problems with segmentation. The main criticisms of the work are largely centered on how the method is evaluated and implemented, rather than fundamental concerns with the validity of the method itself.

Essential revisions:

1) Benchmarks.

All three reviewers had concerns about the nature of the tests for the new method.

Reviewer 1 was particularly concerned that, while a critical advancement of this method is the ability to segment many more regions than previous subcortical atlases, there are still many regions that overlap with existing segmentation tools. Knowing how the reliability of this new approach compares to previous automatic segmentation methods is crucial in being able to know how to trust the overall reliability of the method. The authors should make a direct benchmark against previous methods where they have overlap.

Reviewer 2 thought that the work would certainly benefit from an additional step of out-of-center / out-of-cohort validation analysis. Though they had no serious concern that performance would be unsatisfactory, it would still highlight the extensibility of the method.

Reviewer 3 shared a similar concern, pointing out that automatized methods are usually sensitive to the number of subjects used to build the parcellation, with results from a bigger training cohort being potentially more robust and generalizable. One of the strongest points of the automated method presented in this paper is the adoption of a Bayesian approach, which usually works efficiently for small sample sizes and allows to update previous results when new data comes. Still, it could be highly illustrative to show the performance of the current method depending on the initial training size. From the same set of delineations of the 105 subjects used to test the age bias, what if the authors show the predicted performance from generating the priors on a training set varying its size?

2) Aging analysis.

All three reviewers were confused as to the purpose for and implementation of the aging analysis included in the paper.

Reviewer 1 said that the analysis of the aging effects on the segmentations seemed oddly out of place. It wasn't clear if this is being used to vet the effectiveness of the algorithm (i.e., its ability to pick up on patterns of age-related changes) or the limitations of the algorithm (i.e., the segmentation effectiveness decreases in populations with lower across-voxel contrast). What exactly is the goal with this analysis? Also, why is it limited to only a subset of the regions output from the algorithm?

Reviewer 2 thought that the most important limitation, as acknowledged by the authors, is the bias from anatomical variation through age or disease. The algorithm is shown to be affected by age and most certainly will be affected by contrast and size changes in neurodegenerative disorders. Broader benchmark tests, as proposed above, would likely address this concern.

Reviewer 3 pointed out that, from Figure 4, it is clear how estimated Dice coefficients decrease with age. As it is well noted by the authors, this is likely caused due to the fact that the priors were built from 10 subjects that had an average age of 24.4 years and thus, the highest predicted performance rates are reflected for subjects whose age range (18-40) lies around this average prior age. The authors mentioned in the paper that they plan on modelling the effects of age in the priors in future work. However, they could already address this question in the current work. Since the data used to test this age bias has already been manually delineated, what if the authors generate new priors for this set of delineations, including subjects from all ages, and test whether the predicted Dice coefficients still depend on age, in the same way as was done in Figure 4?

3) Clarity of the algorithm.

Reviewers 1 and 3 had concerns about details of the algorithm itself.

Reviewer 1 thought that, because of the difficulty of the parcellation problem, the algorithm being used is quite complex. The authors do a good job showing the output of each stage of the process (Figure 7 and Figure 8), but it would substantially help general readers to have a schematic of the logic of the algorithm itself.

Reviewer 3 pointed out that it is unclear what is the value for the scale parameter δ that appears in the priors? Is that a free parameter? If so, do results change when this parameter varies? This seems to be a critical aspect of the process (at least insofar as the precision of the results).

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your article "Multi-contrast Anatomical Subcortical Structures Parcellation" for consideration by eLife. Your revised article has been reviewed by three peer reviewers, including Timothy Verstynen as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Michael Frank as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Wolf-Julian Neumann (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). Specifically, we are asking editors to accept without delay manuscripts, like yours, that they judge can stand as eLife papers without additional data, even if they feel that they would make the manuscript stronger. Thus the revisions requested below only address clarity and presentation.

Summary:

The reviewers felt that the revised manuscript is much stronger and focused. There remains only a minor point raised by reviewer #2 that we ask you address.

Essential revisions:

Reviewer #2 (see below) would like to see a more elaborate comparison of the STN segmentation to previous recent methods, given that these sorts of subcortical parcellation methods are likely to be of interest to DBS researchers. While a comparison to the Ewert method (or similar) would be nice, a brief discussion of a subjective comparison of these results would be sufficient to address this concern.

Reviewer #1:

The authors have adequately addressed all of my concerns. Well done.

Reviewer #2:

This is a revised manuscript on a 7T based segmentation approach. My main concern in the primary submission was that at this point the complexity of the data acquisition in combination with the computational approach makes it relatively niche. Other concerns regarded additional validation runs and aging, which the authors have now provided. My feeling is that the authors have made an effort to address the major points raised in the first revision.

There is one minor point that still remains puzzling for me. The fact that the STN is delineated so poorly, when compared to other structures. Given its use as a target in hundreds of thousands of patients for deep brain stimulation, automatized STN segmentation has been validated. The authors cited Ewert et al., which combines a simple normalization with an atlas in MNI space and reached similar and better performance when compared to the presented results here. I would have liked to see the results from this paper reproduced as a comparison here, but if the authors decide not to, I would at least like to invite the authors to discuss why the STN is troublesome for the algorithm and how this could affect use for surgical planning in the future, when 7T becomes available in clinics.

Additionally, I found this a little strange: Figure 3 caption: compared to the most expert of the two human raters.

Reviewer #3:

The authors have addressed all the concerns that I had in the previous version of the manuscript. Important changes in this revision included a benchmark against existing automated parcellation tools, a validation analysis using a test-retest sample from the Human Connectome Project and a thorough examination of training sample size and age biases. All of these changes have significantly increased the quality of the paper and more importantly, provided more clarity and evidence for the benefits of the proposed algorithm for subcortical parcellation. As a consequence, I am more than pleased to recommend the current version of the manuscript for publication.

Essential revisions:

1) Benchmarks.

All three reviewers had concerns about the nature of the tests for the new method.

Reviewer 1 was particularly concerned that, while a critical advancement of this method is the ability to segment many more regions than previous subcortical atlases, there are still many regions that overlap with existing segmentation tools. Knowing how the reliability of this new approach compares to previous automatic segmentation methods is crucial in being able to know how to trust the overall reliability of the method. The authors should make a direct benchmark against previous methods where they have overlap.

We agree with reviewer 1 that comparing to other previously published methods is useful, as few of them provide tables of validation scores. We added a comparison with three popular tools: a multi-atlas registration scheme building directly on our manual delineations, FSL First, and Freesurfer. The results are summarized in a new Table 2, and detailed further in the Materials and methods and Results sections.

In Results:

“Comparison to other automated methods

To provide a basis for comparison, we applied other freely available methods for subcortical structure parcellation to the same ten subjects. MASSP performs similarly or better than Freesurfer, FSL First and a multi-atlas registration using ANTs. Multi-atlas registration provides high accuracy in most structures as well, but is biased toward under-estimating the size of smaller and elongated structures where overlap is systematically reduced across the individual atlas subjects. Multi-atlas registration is also quite computationally intensive when using multiple contrasts at high resolution. Finally, MASSP provides many more structures than Freesurfer and FSL First, and can be easily applied to new structures based on additional manual delineations.”

In Materials and methods:

“Comparisons with other automated methods

To assess the performance of MASSP compared to existing parcellation tools, we ran Freesurfer (Fischl et al., 2002), FSL First (Patenaude et al., 2011) and a multi-atlas registration approach (coregistering 9 of the 10 manually delineated subjects on the remaining one with ANTs (Avants et al., 2008) and labelling each structure by majority voting, similarly to the MAGeT Brain approach of Chakravarty et. al (2013)). Freesurfer and First were run on the skull-stripped R1 map, while the multi-atlas approach used all three R1, R2* and QSM contrasts. All methods were compared in terms of Dice overlap, and we also assessed the presence of a systematic volume bias, defined as the average of the signed difference of the estimated structure volume to the manually delineated volume, normalized by the manually delineated volume.”

Reviewer 2 thought that the work would certainly benefit from an additional step of out-of-center / out-of-cohort validation analysis. Though they had no serious concern that performance would be unsatisfactory, it would still highlight the extensibility of the method.

We thank reviewer 2 for this thought, which led us to define an appropriate methodology to expand the algorithm to other contrasts. While MASSP is based on quantitative MRI, it is not entirely straightforward to simulate the contrast of various MR sequences based on these, as other sources of contrasts and artifacts often influence the imaging. Instead, we took a data-driven approach, coregistering templates with the contrasts of interest and building a statistical mapping of intensities for each defined interface in the MASSP atlas. With this approach, we were able to successfully parcellate data from the Human Connectome Project. We made the following additions to the Materials and methods and Results sections:

In Results:

“Application to new MRI contrasts

Quantitative MRI has only become recently applicable in larger studies, thanks in part to the development of integrated multi-parameter sequences (Weiskopf et al., 2013; Caan et al., 2019). Many data sets, including large-scale open databases, use more common T1- and T2-weighted MRI. In order to test the applicability of MASSP to such contrasts, we obtained the test-retest subset of the Human Connectome Project (HCP, Van Essen et al., 2013) and applied MASSP to the 45 preprocessed and skull-stripped T1- and T2-weighted images from each of the two test and retest sessions. While performing manual delineations on the new contrasts would be preferable, the model is already rich enough to provide stable parcellations. Test-retest reproducibility is similarly high for MASSP and Freesurfer, and are generally in agreement, see Figure 5 and Table 3.”

In Materials and methods:

“Application to new MRI contrasts

Before applying MASSP to unseen contrasts, we need to convert its intensity prior histograms $H_{i|j,R}$ to the new intensities. In order to perform this mapping, we first created a groupwise median of the HCP subjects, by co-registering every subject to the MASSP template using ANTs with nonlinear registration and both T1w, T2w contrasts matched to the template's R1 and R2* maps. […] With these new histograms, we compared the test-retest reliability and overall agreement of MASSP with Freesurfer parcellations included in the HCP preprocessed data set.”

Reviewer 3 shared a similar concern, pointing out that automatized methods are usually sensitive to the number of subjects used to build the parcellation, with results from a bigger training cohort being potentially more robust and generalizable. One of the strongest points of the automated method presented in this paper is the adoption of a Bayesian approach, which usually works efficiently for small sample sizes and allows to update previous results when new data comes. Still, it could be highly illustrative to show the performance of the current method depending on the initial training size. From the same set of delineations of the 105 subjects used to test the age bias, what if the authors show the predicted performance from generating the priors on a training set varying its size?

We thank Reviewer 3 for this suggestion. We used the delineations of the five subcortical structures defined on the entire cohort to define atlases of increasing size, from 3 to 18 subjects. Prior work from us and others (e.g. Bazin et al., 2008; Iglesias et al., 2013) have shown that Bayesian approaches are generally very efficient with regard to the size of the training cohort, and we found indeed that performance stabilized for all structures at 8 subjects. We added the experiment to the manuscript as follows in Results:

“Biases due to atlas size

[…] First, we investigated the impact of atlas size. We randomly assigned half of the subjects from each decade to two groups, and built atlas priors from subsets of 3, 5, 8, 10, 12, 15, and 18 subjects from the first group. The subjects used in the atlas were taken randomly from each decade (18-30, 31-40, 41-50, 51-60, 61-70, 71-80), so as to maximize the age range represented in each atlas. Atlases of increasing size were constructed by adding subjects to previous atlases, so that atlases of increasing complexity include all subjects from simpler atlases. Results applying these atlases to parcellate the second group are given in Figure 6. As in previous studies (Iglesias et al., 2013; Bazin et al., 2008), performance quickly stabilized with atlases of more than five subjects (no significant difference in Welch's t-tests between using 18 subjects or any subset of 8 or more for all structures and measures).”

2) Aging analysis.

All three reviewers were confused as to the purpose for and implementation of the aging analysis included in the paper.

Reviewer 1 said that the analysis of the aging effects on the segmentations seemed oddly out of place. It wasn't clear if this is being used to vet the effectiveness of the algorithm (i.e., its ability to pick up on patterns of age-related changes) or the limitations of the algorithm (i.e., the segmentation effectiveness decreases in populations with lower across-voxel contrast). What exactly is the goal with this analysis? Also, why is it limited to only a subset of the regions output from the algorithm?

Reviewer 2 thought that the most important limitation, as acknowledged by the authors, is the bias from anatomical variation through age or disease. The algorithm is shown to be affected by age and most certainly will be affected by contrast and size changes in neurodegenerative disorders. Broader benchmark tests, as proposed above, would likely address this concern.

Reviewer 3 pointed out that, from Figure 4, it is clear how estimated Dice coefficients decrease with age. As it is well noted by the authors, this is likely caused due to the fact that the priors were built from 10 subjects that had an average age of 24.4 years and thus, the highest predicted performance rates are reflected for subjects whose age range (18-40) lies around this average prior age. The authors mentioned in the paper that they plan on modelling the effects of age in the priors in future work. However, they could already address this question in the current work. Since the data used to test this age bias has already been manually delineated, what if the authors generate new priors for this set of delineations, including subjects from all ages, and test whether the predicted Dice coefficients still depend on age, in the same way as was done in Figure 4?

We thank the reviewers for their detailed comments and suggestions. Indeed, we realize that the aging analysis was not clearly motivated, and that the experiment of Figure 4 was not very informative. Our goal here was to assess the ability of the algorithm to parcellate data from older subjects, a step seldom taken when validating algorithms. In this revision we replaced the experiment of Figure 4 by a more thorough test of biases: using only the five structures for which we have a complete set of delineations, we tested systematically the impact of atlas age range versus testing age group. The results, reproduced below, indicate that while there is a decrease in performance in subjects above 60 year old, the decrease appears largely independent of the choice of the atlas age group. Thus we conclude that the method accuracy does decrease in subjects above 60 year old and likely in patients with neurodegenerative disease. This effect is not merely a reflection of bias in the atlas, but rather the expression of increasing variability in brain MRI. Note however that the following experiment on estimating morphometry and quantitative MRI from automated versus manual parcellations shows a good stability of local measures across age groups, even if structure boundaries and volumes are not perfectly estimated.

In Results:

“Biases due to age differences

To more specifically test the influence of age on parcellation accuracy, we defined six age groups by decade and randomly selected 10 subjects from each group. Each set of subjects was used as priors for the five structures above, and applied to the other age groups. Results are summarized in Figure 7. Examining this age bias, we can see a decrease in performance when parcellating subjects in the range of 60 to 80 years of age. The choice of priors seems to have a limited impact, which varies across structures. Interestingly, using priors from a similar age group is not particularly beneficial.”

3) Clarity of the algorithm.

Reviewers 1 and 3 had concerns about details of the algorithm itself.

Reviewer 1 thought that, because of the difficulty of the parcellation problem, the algorithm being used is quite complex. The authors do a good job showing the output of each stage of the process (Figure 7 and Figure 8), but it would substantially help general readers to have a schematic of the logic of the algorithm itself.

We thank reviewer 1 for this valuable suggestion. We added the following explanatory diagram at the beginning of Results:

MASSP uses a data set of ten expert delineations as a basis for its modeling. From the delineations, an atlas of interfaces between structures, shape skeletons, and interface intensity histograms are generated, and used as prior in a multiple-step non-iterative Bayesian algorithm, see Figure 2 and Materials and methods.

Reviewer 3 pointed out that it is unclear what is the value for the scale parameter δ that appears in the priors? Is that a free parameter? If so, do results change when this parameter varies? This seems to be a critical aspect of the process (at least insofar as the precision of the results).

We thank the reviewer for noticing this omission: δ represents the amount of partial voluming at the interface. In theory it could be tuned (in particular to be increased at smoother boundaries such as the thalamus-internal capsule interface), but setting it to the imaging scale provided the most consistent results across structures. We added this mention to the Materials and methods:

The scale parameter is set to 1 voxel, representing the expected amount of partial voluming.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Essential revisions:

Reviewer #2 (see below) would like to see a more elaborate comparison of the STN segmentation to previous recent methods, given that these sorts of subcortical parcellation methods are likely to be of interest to DBS researchers. While a comparison to the Ewert method (or similar) would be nice, a brief discussion of a subjective comparison of these results would be sufficient to address this concern.

In the article by Ewert et al., the parcellation is performed by co-registration to a very precise high resolution atlas. In their experiment (see their Table 2), they report average Cohen's Kappa of 0.4467 and 0.5240 for the STN and the GPi respectively. We recomputed the corresponding metric for our leave-one-out experiment and obtained 0.4843 and 0.5452 respectively, which is within the same range of accuracy. These structures, the STN particularly, are notoriously difficult to parcellate due to low T1 contrast with the WM of the internal capsule and the proximity of nuclei with similar T2 and T2* contrast (SN, RN GPe). However, the relatively high performance of the atlas co-registration by Ewert et al., 2018 and of multi-atlas co-registration for these structures indicates that further performance improvements may be gained by using a more anatomically precise template. We added the following to the discussion:

“Some enhancement techniques such as building a multi-subject template (Pauli et al., 2018) or adding a denoising step (Bazin et al., 2019) may be beneficial. Co-registration to a high-precision atlas as in (Ewert et al., 2018) may also improve the initial alignment over the MASSP group average template.”

and later on, when discussing applications:

“Important applications of subcortical parcellation also include deep-brain stimulation surgery (Ewert et al., 2018), where the number of structures parcellated by MASSP can help neurosurgeons orient themselves more easily, although precise targeting will still require manual refinements, especially in neurodegenerative diseases.”

Reviewer #2:

[…]

Additionally, I found this a little strange: Figure 3 caption: compared to the most expert of the two human raters.

We changed the text to "compared to the human rater with most neuroanatomical expertise." Here the goal was to compare to our most skilled manual delineations, rather than a consensus between expert and trainee.

Author details

1. Pierre-Louis Bazin

   1. Integrative Model-based Cognitive Neuroscience research unit, University of Amsterdam, Amsterdam, Netherlands
   2. Max-Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany

   Contribution

   Conceptualization, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft

   For correspondence

   pilou.bazin@uva.nl

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-0141-5510

2. Anneke Alkemade

   Integrative Model-based Cognitive Neuroscience research unit, University of Amsterdam, Amsterdam, Netherlands

   Contribution

   Data curation, Formal analysis, Validation, Investigation, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-3234-353X

3. Martijn J Mulder

   1. Integrative Model-based Cognitive Neuroscience research unit, University of Amsterdam, Amsterdam, Netherlands
   2. Psychology Department, Utrecht University, Utrecht, Netherlands

   Contribution

   Resources, Data curation, Software, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

4. Amanda G Henry

   Faculty of Archaeology, Leiden University, Leiden, Netherlands

   Contribution

   Conceptualization, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-2923-4199

5. Birte U Forstmann

   Integrative Model-based Cognitive Neuroscience research unit, University of Amsterdam, Amsterdam, Netherlands

   Contribution

   Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Project administration, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-1005-1675

• 760

  Page views

• 97

  Downloads

• 1

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.