A fast and robust hippocampal subfields segmentation: HSF revealing lifespan volumetric dynamics

Abstract

The hippocampal subfields, pivotal to episodic memory, are distinct both in terms of cyto- and myeloarchitectony. Studying the structure of hippocampal subfields in vivo is crucial to understand volumetric trajectories across the lifespan, from the emergence of episodic memory during early childhood to memory impairments found in older adults. However, segmenting hippocampal subfields on conventional MRI sequences is challenging because of their small size. Furthermore, there is to date no unified segmentation protocol for the hippocampal subfields, which limits comparisons between studies. Therefore, we introduced a novel segmentation tool called HSF short for hippocampal segmentation factory, which leverages an end-to-end deep learning pipeline. First, we validated HSF against currently used tools (ASHS, HIPS, and HippUnfold). Then, we used HSF on 3,750 subjects from the HCP development, young adults, and aging datasets to study the effect of age and sex on hippocampal subfields volumes. Firstly, we showed HSF to be closer to manual segmentation than other currently used tools (p < 0.001), regarding the Dice Coefficient, Hausdorff Distance, and Volumetric Similarity. Then, we showed differential maturation and aging across subfields, with the dentate gyrus being the most affected by age. We also found faster growth and decay in men than in women for most hippocampal subfields. Thus, while we introduced a new, fast and robust end-to-end segmentation tool, our neuroanatomical results concerning the lifespan trajectories of the hippocampal subfields reconcile previous conflicting results.

Keywords: MRI; aging; deep learning; development; semantic segmentation.

Figure 1

Technical description of ROILoc. ROILoc aims at locating and extracting any region of interest on a given MRI.

Figure 2

Complete overview of HSF. A (T1w or T1w) MRI passes through ROILoc to extract the left and right hippocampi. Each subvolume is then randomly augmented to obtain 21 different versions of the same hippocampus. Each segmentation goes through five independent deep learning models, and the final segmentation is a voxel-wise plurality vote across all segmentations. A voxel-wise aleatoric uncertainty map is computed for further post hoc analysis.

Figure 3

Segmentation example from a random subject. The dentate gyrus is in red, CA1/2/3 are in green, yellow, and purple, and the subiculum is in blue.

Figure 4

Cumming estimation plots comparing HSF (T2w) against ASHS (T1w and T2w), HIPS (T1w and T2w), and HippUnfold (HU) (T1w and T2w). The first row illustrates three performance metrics—the dice coefficient (higher is better), the Hausdorff distance (lower is better), and the volumetric similarity (higher is better). The vertical bars in this row represent the mean ±std for each metric group. The dashed line in this row represents the inter-rater reliability for manual segmentation obtained in the earlier study of Bouyeure et al. (2021). As this earlier study only computed the inter-rater comparison as the dice coefficient, it is not available for the other two metrics. The second row depicts the mean effect size (Cohen's d) with a black dot to facilitate statistical comparison between the groups. The black bars in this row represent 95% CIs for variability estimations. The 95% CIs are obtained through non-parametric bootstrap resampling to generate distributions of all possible effect sizes.

Figure 5

Lifespan dynamics of hippocampal subfields. Trend lines (surrounded by standard errors) are defined as natural cubic splines with a number of degrees of freedom minimizing an Akaike Information Criterion. Vertical dashed lines indicate inflection points.

Figure 6

Normalized anteroposterior composition of subfields, going from 0% of the hippocampus (head) to 100% (tail). Vertical black lines are approximate delimiters of the head, body, and tail of the hippocampus.