Alongside this publication, we released our detailed manually defined hippocampal subfields, unsupervised clustering results, topological unfolding framework, Equivolume laminar model solutions, and each of the unfolded morphological and laminar features computed here for the BigBrain dataset. These resources can be used as templates in other studies. Alternatively, registration of these features to new data in our unfolded space can be used to guide future subfield segmentation. In addition, we have also made the code used in this project available via Open Science Framework
(https://osf.io/x542s/). A toolbox for performing hippocampal unfolding, feature extraction, and other useful operations on more general datasets can be found at https://github.com/jordandekraker/Hippunfolding.
3.4.7
Conclusions
In the current project, we mapped the human hippocampus in detail by combining three methods. First, we used a unique dataset, BigBrain, that contains both histological-level detail and macroscopic 3D spatial context. Second, we imposed a topological unfolding framework to the hippocampus. Third, with this framework we extracted a set of
morphological and laminar features, the latter of which have been used prolifically in neocortical characterization and parcellation. Using these methods, we highlight three novel empirical observations. First, unsupervised clustering of these features closely resembles classically defined hippocampal subfields. Secondly, despite traditional reliance on laminar features in histology, morphological features alone are sufficient to closely approximate most hippocampal subfields. Finally, some features such as gyrification in CA1 show, at least qualitatively, subfield-specific anterior-posterior differences that might relate to functional differences described in the extant literature. Overall, these findings highlight new structural characteristics of the hippocampus, and offer promising avenues for improved delineation and characterization of hippocampal subfields using in-vivo neuroimaging.
Chapter 4
4
Hippocampal Automated Topology
This Chapter focuses on the automation and generalization of the hippocampal coordinate framework and feature extraction methods discussed Chapters 2 and 3 to new datasets. This includes the translation of subfield boundaries defined using 3D histology in Chapter 3 to new MRI data via our unfolded (or surface-based) coordinate framework. This framework requires classification of hippocampal tissues into grey matter, high myelin strata (or SRLM), and surrounding structures. In particular, detailed separation of different hippocampal folds or digitations via the SRLM on the inner surface, and
surrounding alveus, white matter, or CSF on the outer surface is required. This is
challenging to do using traditional segmentation methods since these structures are often thin (e.g. as low as 0.3mm in thickness) and their location can be variable depending on the folding configuration of each subject’s hippocampus. To overcome these challenges, we pooled data from previous projects and used it to train a deep neural network (U-Net) for tissue classification. We then applied incremental learning to further train U-Net on the Human Connectome Project (HCP1200) dataset, specifically on 0.7mm3 isotropic T2w images. With post-processing of resulting tissue classifications via a template shape injection method, we were able to apply the unfolded coordinate framework to all
subjects in the HCP1200 dataset. We then applied the subfield boundaries derived in Chapter 3 to all subjects in unfolded space. Collectively, we refer to these methods as Hippocampal Automated Topology (HAT). Features and subfields extracted or applied with these methods were compared to previous work and other extant automated hippocampal subfields segmentation methods. Overall, HAT yielded structural features that agree with previous work and ex-vivo studies, and subfields defined using this method compared favourably to other extant methods in terms of detail and qualitative comparison to anatomical literature.
4.1
Introduction
State-of-the art methods for neocortical parcellation make use of surface-based
quantitative MRI), but also allows similar parcellation schemes to be applied to many subjects despite broad differences in gyral and sulcal patterning. This is challenging to do using more traditional 3D registration based methods where topological breaks often occur when registering subjects with different fine scale gyral or sulcal patterns. A surface-based approach can help overcome this problem by projecting cortical tissue to an unfolded flat or spherical space with 2D topology (see Dale, Fischl and Sereno, 1999; Fischl, Sereno and Dale, 1999; MacDonald et al., 2000; Zijdenbos, Forghani and Evans, 2002; Kim et al., 2005; Glasser et al., 2016; Research and Case Medical Research, 2019). Parcellation performed in this space is constrained from 3D to 2D and can subsequently be projected back to each subject’s native space despite differences in folding patterns. In recent work, we developed a surface-based (or topological) framework for examination of the archicortex, specifically, the hippocampus. As in the neocortex, this enables detailed feature extraction and helps overcome broad inter-individual variability in gyrification, or digitation as it is often referred to in the hippocampus. This is especially critical given recent reports of inter-individual morphological variability in the
hippocampus and its subfields (Ding and Van Hoesen, 2015; Cai et al., 2019; de Flores et
al., 2019; DeKraker et al., 2020). One major challenge with this surface-based approach,
which we aim to address here, is the need for careful manual delineation of tissues that separate the folds or digitations of the hippocampus. This process is costly in time and expertise and reduces the reproducibility and scalability of this method.
Briefly, our surface-based approach to hippocampal feature extraction and parcellation, or subfield delineation, involves indexing hippocampal grey matter on three geodesic axes: anterior-posterior, proximal-distal, and laminar (or inner-outer). The anterior- posterior and proximal-distal indices make up a 2D topologically organized space that can readily be flatmapped. This indexing is determined according to the Laplace equation, where hippocampal grey matter corresponds to the domain and anatomically motivated structures at each edge (e.g. the hippocampal-amygdalar transition area at the anterior terminus) make up boundary conditions. The hippocampal sulcus and
surrounding high myelin laminae (strata radiatum, lacunosum, and moleculaire or SRLM) separate the inward curling of the hippocampus, as well as the inner boundaries between digitations, while the alveus and surrounding structures (including the third ventricle and
inferior longitudinal fasciculus) separate the outer boundaries between digitations. Topological breaks or bridges between digitations in this grey matter label can introduce major distortions to the Laplacian solution, and so highly detailed segmentation of these structures is required.
Artificial neural networks, and particularly the U-Net architecture (Chen et al., 2018) are becoming increasingly popular in medical image segmentation. U-Net can take advantage of local and global image features in 3D, and in principle it could also discover 2D
topological structure in 3D data or leverage subtle differences in thickness or intensity between hippocampal subfields. However, there is no evidence that current applications of U-Net for hippocampal and hippocampal subfield segmentation leverage such
sophisticated topological modelling or subtle subfield-related intensity differences. Instead, results closely resemble the manually labelled data which they were trained on, which does not show detailed digitations and corresponding topological shifts in
subfields (Shi, Cheng and Liu, 2019; Zhu et al., 2019; Yang et al., 2020). This highlights a need for more elaborate methods and/or more detailed training data.
In the current work, we apply U-Net segmentation in concert with our previously developed hippocampal unfolding framework, which we jointly refer to here as Hippocampal Automated Topology or HAT. Starting with detailed hippocampal
segmentations from a collective of previous studies, we aimed to develop an automated method for obtaining detailed segmentation of hippocampal tissue classes in the Human Connectome Project (HCP1200) dataset (Glasser et al., 2013). These tissue classes can then be used to unfold hippocampal grey matter under our previously developed hippocampal coordinate framework (DeKraker et al., 2018). That is, we use U-Net to segment grey and white matter tissue classes within the hippocampus (most critically hippocampal grey matter and the high myelin SRLM, with background tissues spanning CSF, surrounding white matter, and alveus). Following post-processing, we then apply our previously developed unfolded coordinate framework and define hippocampal subfields according to highly detailed boundaries derived from 3D histology in previous work (DeKraker et al., 2020). It should be noted that while the same boundaries are applied to all hippocampi in unfolded space, they may vary considerably between
subjects in native space depending on the folding configuration (particularly the number and prominence of digitations) in each subject's hippocampus. This entire pipeline is overviewed in Figure 19.
Figure 19. Overview of full Hippocampal Automated Topology pipeline. Step 1: left and right hippocampi are cropped and resampled to 0.3mm3 isotropic obliquely to
the hippocampal long-axis. Step 2: images are segmented via U-Net architecture and then post-processed using template shape injection. Step 3: Previously developed Laplace coordinate framework is applied to the domain of hippocampal grey matter tissue. Step 4: A single subfield atlas defined in 2D unfolded space, in this case generated from BigBrain 3D histology, is propagated to a given subject’s native hippocampal folding configuration. Shown here is the example of the right hippocampus of HCP1200 subject 108020.