Multi-contrast submillimetric 3 Tesla hippocampal subfield segmentation protocol and dataset

The hippocampus is composed of distinct anatomical subregions that participate in multiple cognitive processes and are differentially affected in prevalent neurological and psychiatric conditions. Advances in high-field MRI allow for the non-invasive identification of hippocampal substructure. These approaches, however, demand time-consuming manual segmentation that relies heavily on anatomical expertise. Here, we share manual labels and associated high-resolution MRI data (MNI-HISUB25; submillimetric T1- and T2-weighted images, detailed sequence information, and stereotaxic probabilistic anatomical maps) based on 25 healthy subjects. Data were acquired on a widely available 3 Tesla MRI system using a 32 phased-array head coil. The protocol divided the hippocampal formation into three subregions: subicular complex, merged Cornu Ammonis 1, 2 and 3 (CA1-3) subfields, and CA4-dentate gyrus (CA4-DG). Segmentation was guided by consistent intensity and morphology characteristics of the densely myelinated molecular layer together with few geometry-based boundaries flexible to overall mesiotemporal anatomy, and achieved excellent intra-/inter-rater reliability (Dice index ≥90/87%). The dataset can inform neuroimaging assessments of the mesiotemporal lobe and help to develop segmentation algorithms relevant for basic and clinical neurosciences.

The hippocampus is composed of distinct anatomical subregions that participate in multiple cognitive processes and are differentially affected in prevalent neurological and psychiatric conditions. Advances in high-field MRI allow for the non-invasive identification of hippocampal substructure. These approaches, however, demand time-consuming manual segmentation that relies heavily on anatomical expertise. Here, we share manual labels and associated high-resolution MRI data (MNI-HISUB25; submillimetric T1-and T2-weighted images, detailed sequence information, and stereotaxic probabilistic anatomical maps) based on 25 healthy subjects. Data were acquired on a widely available 3 Tesla MRI system using a 32 phasedarray head coil. The protocol divided the hippocampal formation into three subregions: subicular complex, merged Cornu Ammonis 1, 2 and 3 (CA1-3) subfields, and CA4-dentate gyrus (CA4-DG). Segmentation was guided by consistent intensity and morphology characteristics of the densely myelinated molecular layer together with few geometry-based boundaries flexible to overall mesiotemporal anatomy, and achieved excellent intra-/inter-rater reliability (Dice index ≥90/87%). The dataset can inform neuroimaging assessments of the mesiotemporal lobe and help to develop segmentation algorithms relevant for basic and clinical neurosciences.

Background & Summary
The hippocampus has been a focus of neuroscience research for decades. Highly segregated connectional properties have promoted its use as a model system. The hippocampus plays an important role in multiple cognitive processes, particularly declarative memory 1,2 ; its structural compromise is a hallmark of prevalent neurological and psychiatric disorders, such as temporal lobe epilepsy 3 , Alzheimer's disease 4,5 , depression 6 , and schizophrenia 7 .
Prior to the advent of sophisticated histological staining techniques 8 , the hippocampal formation was described as a single entity despite its complex histo-morphology. Since the description by Ramon y Cajal 9 , several histological subdivisions have been proposed [10][11][12] . Similarly, neuroimaging studies have generally considered the hippocampus as a single structure, constrained by limited spatial resolution 13 . Developments in high-field MRI at 3 Tesla and beyond, together with the use of phased-array head coils, offer new opportunities to appraise its internal structure by unveiling strata rich in white matter, and improved identification of the hippocampal sulcus, which separates Cornu Ammonis (CA) and subiculum from the dentate gyrus (DG). Paralleling advances in hardware, a number of studies have provided MRI-based guidelines to manually segment hippocampal subfields [14][15][16][17][18][19][20][21][22][23] . While substantial progress has been made, challenges remain, particularly when attempting to separate individual CA subfields from one another, which compromises reliability within and across analysts. From a practical perspective, manual segmentations require anatomical expertise and are often prohibitively time-consuming.
Here, we share a dataset containing manual segmentations of hippocampal subfields together with submillimetric multi-spectral images in 25 healthy individuals. To facilitate local implementation and independent verification, we share detailed MR sequence information as well; importantly, all data were acquired in a clinically-feasible scan time on a widely available 3 Tesla MRI system.
Opting for high reliability, segmentations were based on a protocol that divided the hippocampal formation into consistently identifiable subregions, guided by intensity and morphology of the densely myelinated molecular layer, together with few geometry-based boundaries flexible to overall mesiotemporal anatomy. Specifically, we combined presubiculum, parasubiculum, and subiculum proper into a single label (subiculum), joined CA1, 2, and 3 (CA1-3), and merged CA4 with the DG (CA4-DG). While segmentation relied primarily on T1-weighted (T1w) data, T2-weighted (T2w) images offered additional guidance. We provide the full set of multispectral images in high-resolution native and stereotaxic (MNI152) space, the manual labels, together with a probabilistic atlas that can inform functional and structural imaging assessments of the hippocampal formation. Moreover, our datasets can be used to develop new protocols, validate existing ones and design automated algorithms relevant for basic as well as clinical neurosciences.

Participants
We studied 25 healthy individuals (12 males; 21-53 years, mean ± s.d. age = 31.2 ± 7.5 years; Table 1), recruited through advertisement. All participants had normal or corrected-to-normal vision; none of them suffered from neurological, psychiatric, or somatic diseases. The Ethics Committee of the Montreal Neurological Institute and Hospital approved the study and written informed consent was obtained from all participants in accordance with the standards of the Declaration of Helsinki. Participants gave their written informed consent prior to scanning and received a monetary compensation.

Scan parameters
MRI data were acquired on a 3 Tesla Siemens TimTrio scanner using a 32-channel head coil. We obtained two sets of T1w images: a 3D magnetization-prepared rapid-acquisition gradient echo (

Pre-processing
MRI data files were converted from DICOM to MINC (*.mnc) format using dcm2mnc with dicom header anonymization. Images underwent automated correction for intensity non-uniformity and intensity standardization 24 . Millimetric and submillimetric T1w MRI volumes were linearly registered to the high-resolution MNI-ICBM152 template 25,26 . T2w images were linearly registered to the millimetric T1w MRI in native space; the resulting transformation matrix was concatenated with the matrix that mapped the millimetric T1w image to the MNI space, thereby linearly mapping T2w images to this template. During the final registration of submillimetric T1w and T2w data to MNI space, images were resampled to a resolution of 0.4 × 0.4 × 0.4 mm 3 , yielding a voxel volume of 0.064 mm 3 . To reduce interpolation artifacts given the anisotropic resolution of the T2w data, we applied a non-local up-sampling method that recovers high frequency information using a data-adaptive patch-based reconstruction together with a subsampling coherence constraint 27 . MNI-space structural scans were subsequently anonymized by zeroing out the voxels in the vicinity of the facial surface, teeth, and auricles following a previously described procedure 28 . For data sharing, images were converted to NIfTI (*.nii) format using mnc2nii. Please see Fig. 1 for a schematic overview of the preprocessing steps and data quality.

Protocol description
A single rater (JKY), blinded to case identities, carried out all segmentations using a 3D viewer (http:// www.bic.mni.mcgill.ca/ServicesSoftwareVisualization/). Subfield segmentation took approximately 16 h per individual (8 h per hemisphere). Boundaries were based on anatomical descriptions of the hippocampus by Duvernoy 29 and Insausti 30 . As spatial relationships between subfields vary along the hippocampal long axis, landmarks are separately described for the hippocampal head ( Fig. 2a-e), body (Fig. 2f), and tail ( Fig. 2g-j). These segments were defined as in our previous protocol 31 .
Segmentations were primarily performed on coronal T1w images, with cross-referencing to sagittal/ axial views. T2w data eased the identification of the densely myelinated and thick molecular layer of the subiculum (forming its superior border). This layer is hyperintense on T1w and hypointense on T2w images ( Fig. 2b-i, k); it is contiguous with, but distinct from the thinner molecular layer of CA1 (ref. 30). The second landmark is the molecular layer of the DG and that of CA fused across the vestigial hippocampal sulcus; this ensemble is visible as a T1w-hyperintense/T2w-hypointense band (Fig. 2c-i). The molecular layers, along with residual vascular cavities that follow the sulcal route, consistently appear on T2w images and separate the DG from the subiculum (inferiorly and medially) and the CA (inferiorly, laterally, and superiorly). We included alveus and fimbria in the CA1-3 label.
a) Hippocampal head. The hippocampal head includes the subiculum, CA1-3, and small portions of the DG. Its rostral-most section is composed of the subiculum only 30 (Fig. 2a) the subiculum, separating it from the overlying amygdala; cross-referencing to the sagittal view confirmed this boundary. The inferior subicular boundary is formed by parahippocampal white matter running along the entire rostro-caudal extent of the hippocampus. Perforant projections from the entorhinal cortex to the subiculum occasionally blurred this boundary; in this case, we identified the subiculum by cross-referencing to axial/sagittal views. As the exact boundary between subiculum and infero-medial entorhinal cortex cannot be visualized on MRI, it was defined by extending a straight line along the graywhite matter border at the crown of the parahippocampal gyrus until it reached the cerebro-spinal fluid in the ambient cistern 32 .
When CA1 first becomes visible, it runs parallel to the subiculum; for a few slices, the molecular layer of the subiculum separates both structures, with CA1-3 on the top (Fig. 2b). More posteriorly, given the overlap (rather than sharp transition) between the pyramidal layers of CA1 and subiculum 30 , we drew a line along the hippocampal sulcus pointing towards the fundus of the collateral sulcus (Fig. 2b,c). This often-oblique line has been previously used to describe this boundary 19 .
The hippocampal head exhibits 3-4 digitations before turning medially to form the posterior uncus. Each digitation encapsulates an extension of the DG. At the level of the head, however, the DG molecular layer that would have allowed for its identification cannot be visualized. For consistency, we merged CA   and DG at this level (Fig. 2c). We could reliably segment CA4-DG at the junction of head and body, where the medial surface of the DG (known as margo denticulatus) becomes visible (Fig. 2d).
b) Hippocampal body. Head and body interface at the caudal end of the uncus 31 (Fig. 2e,f). Here, the margo denticulatus of the DG has a characteristic toothed appearance and is separated from the overhanging fimbria by the fimbriodentate sulcus. Coronally, the orientation of the hippocampal body varies along its rostro-caudal direction both across and within individuals. The term malrotation has been coined to describe this abnormal shifting/rotations of the long hippocampal axis relative to the horizontal plane 33,34 , which likely affects the relative boundary between subiculum and CA1. To determine this border, we adapted our guidelines based on the position of the hippocampus on coronal slices: (1) if the left hippocampus was oriented counter-clockwise (clockwise for the right hippocampus), the boundary was defined as the extension of the line corresponding to the superior subicular border (Fig. 2e); (2) if the hippocampus was horizontally positioned, the border was defined as a line drawn from the lateral-most point of the subicular molecular layer at a 45 degrees angle until it reached the underlying white matter (Fig. 2f); (3) if the left hippocampus was oriented clockwise (counter-clockwise for the right), the border followed a line drawn from the lateral-most point of the subicular molecular layer towards the fundus of the collateral sulcus (Fig. 2d). Inferior and medial boundaries of the subiculum remained the same as in the head.
CA and DG form two U-shaped interlocking laminae, one fitting into the other, and separated from each other by the hippocampal sulcus. For consistency, voxels corresponding to the fused molecular layers of the CA1-3 and DG were assigned to CA4-DG. As the CA3-CA4 boundary cannot be resolved on MRI, the superior border of CA4-DG was drawn as the horizontal continuation of the hippocampal sulcus, from its most medially visible point towards the fimbriodentate sulcus.
c) Hippocampal tail. The junction between body and tail was set as the rostral-most slice at which the crus fornix becomes fully visible (Fig. 2g) 31 . In the initial segment of the tail, the CA1-subiculum boundary was determined to be the infero-lateral extension of the superior subicular border (Fig. 2g,h). Inferior and medial borders of the subiculum were defined as in the body. In the initial portion of the tail, CA1 is deeply located, hidden by the subiculum; more posteriorly, it appears at the surface of the parahippocampal gyrus, progressively replacing the subiculum. The exact posterior subicular border is not visible on MRI: we consistently chose it to be the posterior-most coronal slice on which the thalamus could be seen (Fig. 2h,i), verified on sagittal view. We excluded the isthmus of the cingulate gyrus, which replaces the subiculum in the middle and terminal segments of the tail, by excluding grey matter inferior to the hippocampal sulcus, best visualized sagittally. The hippocampal sulcus separates the DG from the subiculum in the initial segment, and from CA1-3 in the initial and middle segments. Furthermore, the fused molecular layers of CA and DG allowed us to visualize the caudal border of the DG on the sagittal view.
The posterior hippocampal end belongs to CA1-3 (Fig. 2j) and faces the cerebrospinal fluid of the lateral ventricle medially and of the atrium laterally. This boundary was best seen sagittaly (Fig. 2k). While fimbria and alveus were included in CA1-3, we excluded the crus fornix (Fig. 2g). The latter joins the splenium of the corpus callosum.

Data Records
The submillimetric 3 Tesla dataset are highly suitable for the development and cross-validation of future manual or automatic segmentation protocols. MRI data and subfield segmentations of all participants, detailed scan parameters, as well as stereotaxic probabilistic maps are available on Dryad (Data Citation 1) and NITRC (Data Citation 2). A README file with a detailed description of the content of all downloads is available there as well. MRI data files were converted from to DICOM to MINC format (using dcm2mnc) before processing, and to NIfTI (using mnc2nii) after processing. For every subject, high-resolution T1w and T2w data are available in 0.4 mm isotropic MNI152 space as well as in their native spaces. For registration purposes, the 1 × 1 × 1 mm 3 T1w data is also provided in native and stereotaxic space. Labels in NIfTI format of the subiculum, CA1-3 and CA4-DG are provided in the highresolution MNI152 space. We furthermore provide probabilistic anatomical maps of each subfield in 1 × 1 × 1 mm 3 MNI152 space. To anonymize data, centre-specific study and participant codes have been removed using an automated procedure. MRI data have been de-faced. All participants were given sequential integer IDs with an 'S' prefix.

Technical Validation
Contrast-to-noise ratio To obtain a quantitative index of MRI data quality, we estimated Contrast-to-Noise ratio (CNR), similar to the approach carried out in a recently published study 21 . In short, an eroded mask of the CA1-3 was compared with an equivalently-sized mask of the temporal lobe white matter inferior to it. The CNR was estimated using the following formula:  Table 3. Intra-and inter-rater reliability assessment. Intra-and inter-rater reliability JKY segmented subfields of 10 hippocampi (5 left, 5 right) from 10 different subjects twice, 6 months apart. We assessed inter-rater reliability comparing subfield delineations of 10 hippocampi segmented by JKY and another observer (KL), blinded to each other's segmentation. Reliability was quantified using Dice overlap indices between two labels 35 , D = 2 × |M 1 ∩ M 2 |/(|M 1 |+|M 2 |) × 100%, where M 1 is the 1st label, M 2 the 2nd label; M 1 ∩ M 2 is the intersection of M 1 and M 2 ; |.| is the volume operator. We also calculated intra-class correlations (ICC). The Dice index quantifies the overlap of two labels geometrically, whereas ICC calculates statistical similarity. To approximate the actual distribution of reliability values, we employed 1,000 bootstrap-based subsamplings and computed 95% confidence intervals. Table 3 displays mean ± s.d. as well as bootstrap confidence interval of Dice indices for individual subfileds. Overall, indices were ≥90 and 87% for intra-and inter-rater reliability, respectively. The ICC ranged from 0.91 to 0.96 within and from 0.73 to 0.91 between raters.

Probabilistic anatomical maps
For each MNI152-space subfield label, we generated statistical anatomical maps that outline the probability of subfield location across participants (Fig. 3).

Usage Notes
The procedures we employed in this study resulted in a high-resolution 3 Tesla dataset containing submillimetric MRI data in native and MNI152 space, together with manual labels of three hippocampal subfields in MNI152 space. Data are shared in documented standard formats, such as NIfTI or plain text files, to enable further processing in arbitrary analysis environments with no imposed dependencies on proprietary tools. Exam card printouts from the scanner are also available for local implementation of the image acquisition protocol. All processing performed on the released data article were produced by openly accessible software on standard computer workstations. Data are available on a curated open access repository (Data Citation 1) and on NIRTC (Data Citation 2).