Manual delineation approaches for direct imaging of the subcortex

Anatomical atlases are essential for directing brain research, interpreting and standardizing results, as well as guiding surgery (Devlin and Poldrack 2007; Forstmann et al. 2017). Particularly atlases of the human subcortex continue to gain interest in view of the growing number of structures that could potentially serve as targets for deep brain stimulation surgery (DBS) (Lozano et al. 2019). The golden standard approach for the identification and parcellation of brain structures is the manual delineation on individual magnetic resonance imaging (MRI) contrasts in their native space. This allows the subsequent creation of probabilistic atlases, which capture the interindividual anatomical variation in the brain, via registration to standard MRI-space (Mazziotta et al. 1995).

Creating 3-dimensional (3D) delineations of the approximately 455 subcortical structures using individual MRI contrasts represents a major feat (Alkemade et al. 2013). To perform manual delineation in a consistent and reproducible manner clear anatomical descriptions are indispensable. In view of the many major and minor anatomical differences between people, flexible knowledge of the human brain is crucial for performing delineations. It can be challenging to mentally image the 3D structure together with the relative orientation of individual brain nuclei on MRI contrasts using a histological atlas such as those published by Schaltenbrand and Wahren (1977) and Mai et al. (2015). Familiarization with the 3D structure and relative orientation of individual subcortical brain structures will, therefore, contribute to the development of the flexible knowledge on human neuroanatomy, which is crucial to MRI parcellation efforts.

To facilitate delineation efforts on the human subcortex, we created a series of illustrated parcellation protocols which can be reused to identify and outline a set of 21 individual structures. Each protocol starts with a post mortem blockface image derived from our previous studies for anatomical orientation (Alkemade et al. 2020b). The list of 21 structures is by far not exhaustive and can viewed as a first step that can be extended in the future. Structures included were selected based on their visibility, their potential role as a target structure for deep brain stimulation surgery, or their practical use as an anatomical landmark. We would like to note that these structures include aggregations of anatomical substructures such as the thalamus (Tha), as well as deeper brain structures that are not of subcortical origin, such as the amygdala, and the ventricular system. This work is not intended to determine whether brain structures belong to the cortex or subcortex, but to share delineation protocols that can be reused for the delineation of individual brain structures. 7 Tesla (T) MRI scans, with submillimeter resolution were used to create these descriptions, which has resulted in visual representations of the delineated structures in all anatomical planes (Alkemade et al. 2020a). These manual delineations have also formed the basis for training of the multi-contrast anatomical subcortical structures parcellation (MASSP) algorithm which has been published previously (Bazin et al. 2020). To improve flexible knowledge, we present individual delineation instructions on three separate datasets. The delineation protocols were tested by creation of manual delineations by two independent raters on a training dataset of ten young healthy individuals. Indicators of interrater agreement were used as a measure for the reliability of the resulting parcellations, and the usability of the protocols.

All structures were parcellated in FSLview or FSLeyes (fsl.fmrib.ox.ac.uk) using MRI contrasts obtained from a subset of ten healthy volunteers (age 20–49, eight females) from the Amsterdam ultra-high field adult lifespan database (AHEAD) described previously (Alkemade et al. 2020a). All participants were healthy as assessed through self-reports.

Available contrasts included the following: T1-weighted, T1- and R1-maps, T2*-maps, and quantitative susceptibility mapping. Contrasts for the delineations of individual borders were determined experimentally and per structure. Raters varied across structures and included both experienced and novel raters. Novel raters first performed independent test parcellations on which feedback was provided by an experienced anatomist (AA) on additional available scans. Parcellations were performed under standardized conditions in individual space, in which the order of participants as well as the first hemisphere to delineate was randomized. Lighting circumstances were kept stable, and raters had access to high-resolution computer screens and drawing pads for parcellation purposes.

Contrasts were calculated from a single MRI acquisition using the MP2RAGEME sequence described previously (Caan et al. 2019). The MP2RAGEME sequence allows the calculation of T1 and T2* derived contrasts from a single scan acquisition, resulting in intrinsic alignment of the resulting contrasts.

Dice similarity coefficients and dilated Dice coefficients were calculated based on manual delineations of two independent raters, as described previously (Dice and Dice 1945; Bazin et al. 2016, 2020; Alkemade et al. 2020a; Trutti et al. 2021). Protocol layouts were then standardized to create a coherent series of guidelines to delineate 21 Individual structures (see Table 1). Parcellation procedures are presented below and are illustrated using screen shots of the resulting masks in the coronal, axial, and sagittal plane. All procedures contain a concise description of the structure and an image from a human brain cut in serial coronal sections for anatomical orientation (Alkemade et al. 2020b). Additionally, a concise description of the procedure and illustrations of the parcellations at various anatomical levels on T1-weighted and quantitative susceptibility mapping (QSM) contrasts are included. Table 1 indicates the MRI contrast used for the parcellation procedure for each individual structure. We would like to note that the visibility of individual anatomical borders varies across contrasts. This potentially hampers the reuse of the protocols for parcellations on clinical scans, in which not all MR contrasts presented here are included. We share manual parcellations from single raters together with the MRI contrasts of three individual participants. This allows users to inspect the data in detail in the coronal, axial and sagittal plane, on T1-weighted, T1-maps, T2*-maps, R1-maps, R2*-maps, and QSM, and help to determine their usability on other datasets including clinical scans acquired using lower field strengths.

Manual parcellations were performed by two individual raters for each structure on ten individual participants. Structures crossing the midline were not separated into observations for left and right hemispheres. In Table 1 Dice similarity coefficients and dilated Dice coefficients are presented.

The human subcortex comprises a wealth of individual brain structures that are only sparsely mapped and included in commonly used brain atlases for MRI (Alkemade et al. 2013). Here, we present delineation protocols of 21 of these structures (Table 1), together with our results supporting the replicability of the masks when using these protocols. The illustrations presented in the protocols are not exhaustive. We have incorporated slice by slice information through the sharing of three individual datasets, which provides a more informative alternative as it shows the parcellations in the coronal, axial, and sagittal planes (see data availability). Decisions to include or exclude voxels vary between slices when following a slice by slice procedure as our protocols describe. This can result in a jagged appearance of individual structure. Smoothing of these edges would require additional arbitrary decisions, and are unlikely to substantially influence the parcellation results other than their jagged appearance, nor the use of the parcellations in further analyses. We therefore decided not to smooth the parcellation results.

With these descriptions, we share a methodology that can be reused by other research groups for manual parcellation purposes, including validation of automated delineation procedures in comparable as well as different MRI contrasts. We confirmed that these protocols were intelligible for both trained and less experienced anatomists, which performed the parcellations through calculation of Dice similarity, and dilated Dice similarity coefficients. The typical reported range of a good intra-rater reliability is higher than 0.75 as reflected by the Dice similarity coefficient (Dice and Dice 1945). We observed that structures with a limited diameter in any direction and/or delineated based on landmarks, such as the VTA, SCG and the PPN, tended to have somewhat lower Dice similarity coefficients as compared to those with borders defined based on image contrast, as well as larger structures. For such structures the Dice similarity coefficient may result in an underestimation of the interrater agreement. The dilated Dice score is a potentially more representative measure for the agreement between raters regardless of size, as the smaller structures can reach high levels of overlap (Bazin et al. 2016, 2020; Trutti et al. 2021). The dilated Dice score (0.82–1.00, Table 1), which allows a single voxel of uncertainty indicates that the observed variation results mainly from differences in border detection, and not structural differences in the parcellation approaches of the individual raters. In some cases, borders between two structures may be difficult to detect, in part due to partial voluming effects. This may result in the attribution of some voxels to more than one individual structure. The use of conjunct masks reduces this problem.

We would like to emphasize that we developed and validated these protocols on data of the AHEAD, which is a 7T anatomical dataset, acquired with submillimeter resolution (Alkemade et al. 2020a). The question, therefore, rises to what extent the presented protocols can be reused for clinical datasets. We have previously published on the parcellation of the GPi/e, SN, RN and STN in ten participants of the AHEAD, which were also scanned in a 3T scanner using a clinical scan and an optimized scan. The results confirmed reliable parcellation in optimized scans with a 1-mm isotropic voxel resolution. However, manual parcellations did not yield an anatomically plausible shape of the STN using the clinical scan. This was attributed to the use anisotropic voxels (Mulder et al. 2019; Isaacs et al. 2020). With the approval of 7T systems for clinical practice, the use of UHF submillimeter imaging for the identification of DBS targets in the clinic now comes within reach.

The parcellation protocols presented, in some cases, led to the inclusion of (parts of) adjacent structures, or the omission of part of the structure of interest. These arbitrary choices were included in the parcellation protocols for practical reasons and in most cases were guided by a lack of anatomical contrast, which hampered a more precise border determination. Ongoing developments in MRI hardware and scan acquisition protocols are expected to improve anatomical contrast even further, which may warrant updating of the presented descriptions accordingly.

Automated anatomical labeling procedures provide the promise of reliable identification and parcellation of structures of interest, although anatomical validation is required to confirm the correct identification of the anatomy (Poldrack 2007). The presented parcellation procedures have formed the basis for the training and validation of the MASSP algorithm (Bazin et al. 2020) and thus can be reused by researchers aiming to operationalize the use of MASSP on different MRI datasets, allowing them to apply the same level of rigor as we did in the validation of the algorithm for use on the 7T MP2RAGEME contrasts as described by Caan et al. (2019). With adaptions these protocols are expected to be of value as well for their use on different MRI contrasts due to the general knowledge they provide on the shape and size as well as the spatial orientation of and between the structures in the descriptions. The development of automated anatomical labeling procedures represents a growing field of research. These efforts are invaluable for the analyses of large neuroimaging datasets such as the human connectome project (Van Essen et al. 2012; Glasser et al. 2016), which in view of the laborious nature of manual parcellation does not lend itself for manual assessments of individual neuroanatomy. Automated procedures are, therefore, indispensable and require reproducible validation strategies to ensure the quality of automated labeling procedures performed in datasets with slightly different contrast characteristics.

A major challenge is to capture the substantial anatomical variability that is observed between participants. Variability occurs not only in shape, size, and location, but also in iron content, and MRI characteristics as a result of aging and disease (Draganski et al. 2011; Fjell et al. 2013; Keuken et al. 2013, 2017; Andersen et al. 2014; Yeatman et al. 2014; Alkemade et al. 2017; Zhang et al. 2018; Herting et al. 2018; Hill et al. 2018; Turner 2019). It is possible that this variability contributes to differences in parcellation results obtained using both manual and automated parcellation procedures. Validation of both the automated and manual parcellation procedures is, therefore, warranted when applying existing methods on participant cohorts that differ in scan acquisition parameters, and/or aging and disease status. This validation can be achieved by determining Dice similarity coefficients and dilated Dice coefficients between manual and automated parcellation results.

Comparison of anatomical variation can broaden our understanding the normal and aberrant brain anatomy, including healthy aging as well as the development of atrophy as a result of neurodegenerative disease. In general, high-quality training data will contribute to a high performance of the algorithm implemented for automated parcellation. The use of standardized manual parcellation protocols will, therefore, contribute to the performance of automated parcellation procedures.