Analyzing Regional Organization of the Human Hippocampus in 3D-PLI Using Contrastive Learning and Geometric Unfolding

TL;DR

This work demonstrates a novel method to analyze the regional organization of the human hippocampus in 3D-PLI by combining recent advances in unfolding methods with deep texture features obtained using a self-supervised contrastive learning approach.

Abstract

Understanding the cortical organization of the human brain requires interpretable descriptors for distinct structural and functional imaging data. 3D polarized light imaging (3D-PLI) is an imaging modality for visualizing fiber architecture in postmortem brains with high resolution that also captures the presence of cell bodies, for example, to identify hippocampal subfields. The rich texture in 3D-PLI images, however, makes this modality particularly difficult to analyze and best practices for characterizing architectonic patterns still need to be established. In this work, we demonstrate a novel method to analyze the regional organization of the human hippocampus in 3D-PLI by combining recent advances in unfolding methods with deep texture features obtained using a self-supervised contrastive learning approach. We identify clusters in the representations that correspond well with classical descriptions of hippocampal subfields, lending validity to the developed methodology.

Figures (3)

• Figure 1: 3D-PLI measurement process and example parameter maps for a brain section through a human hippocampus. (A) Setup of the polarizing microscope (PM). For rotations $\rho$ of a linear polarizer (i), an intensity profile is recorded for each pixel (ii), from which transmittance $I_T$, fiber direction $\varphi$ and inclination $\alpha$ (from retardation $r$) can be derived. $\varphi$ and $\alpha$ determine the 3D orientation of nerve fibers (iii). (B) Example transmittance map $I_T$ with hippocampal subfield labels CA1, CA2, CA3, CA4 and the Subicular complex. (C) Fiber orientation map (FOM) of the same section as in (B).
• Figure 2: Overview of the deep texture feature extraction and unfolding. (A) Encoder models $f$ are trained by (i) sampling positive pairs at a fixed distance $r$ either in-plane (CL-2D) or across brain sections (CL-3D) in (ii) a contrastive learning framework. (B) Selected feature maps generated by CL-3D. A sliding window approach is used to generate feature maps for whole brain sections. Feature maps are overlaid with transmittance maps for reference. (C) Coronal view of surfaces extracted using HippUnfold, overlaid on feature #143 and transmittance (i). Deep texture features by CL-3D are sampled along interpolated vertices between geometrical inner and outer surfaces (green arrows; ii) and concatenated into single vectors. (D) Scree plot showing highest eigenvalues for the first 6 PCA components calculated for the concatenated texture features. (E) Unfolded projections of the concatenated features by CL-3D onto the PCA components with largest explained variance. Components 1-6 are shown on the smoothed mid-surface (left) and in unfolded space (right). Gray marks missing data.
• Figure 3: Comparison of k-means clustering for 6 clusters with anatomically identified subfield labels (A). Clusters of CL-3D texture features (B) more closely resemble the subfield labels than clusters of mean transmittance $\bar{I}_T$ and fractional anisotropy FA (C). Images show the midsurface in folded (left) and unfolded space (right). Black contours and white spots show positions of missing data.