Automated hippocampal unfolding for morphometry and subfield segmentation with HippUnfold

  1. Jordan DeKraker  Is a corresponding author
  2. Roy AM Haast
  3. Mohamed D Yousif
  4. Bradley Karat
  5. Jonathan C Lau
  6. Stefan Köhler
  7. Ali R Khan  Is a corresponding author
  1. Robarts Research Institute, Schulich School of Medicine and Dentistry, The University of Western Ontario, Canada
  2. Western Institute for Neuroscience, The University of Western Ontario, Canada
  3. Department of Clinical Neurological Sciences, Division of Neurosurgery, Schulich School of Medicine & Dentistry, The University of Western Ontario, Canada
  4. School of Biomedical Engineering, The University of Western Ontario, Canada
  5. Department of Psychology, Faculty of Social Science, The University of Western Ontario, Canada
  6. Department of Medical Biophysics, Schulich School of Medicine and Dentistry, The University of Western Ontario, Canada
7 figures, 1 table and 5 additional files

Figures

Figure 1 with 1 supplement
Overview of HippUnfold pipeline.

First, input MRI images are preprocessed and cropped around the left and right hippocampi. Second, a U-Net neural network architecture (nnUNet; Isensee et al., 2021) is used to segment hippocampal …

Figure 1—figure supplement 1
Diagram of the nnU-net architecture used for HippUnfold.

This architecture was automatically configured as the 3D_fullres network, using the 128 × 256 × 128 (0.3 × 0.3 × 0.3 mm3) hippocampal subregion images as training data. All conv3D blocks have stride …

Figure 2 with 1 supplement
Average hippocampal folded and unfolded surfaces showing subfields, morphometric, and quantitative MRI measures from the Human Connectome Project-Young Adult (HCP-YA) test dataset (see Table 1 of Materials and methods).

The same topologically defined subfields were applied in unfolded space to all subjects (top), which are also overlaid on quantitative MRI plots (black lines). The dentate gyrus (DG) is represented …

Figure 2—figure supplement 1
Examination of distortions (or difference in vertex spacing) between an average folded and unfolded space.

Distortions were greatest in the tail of the hippocampus where its proximal-distal distance becomes quite narrow.

Figure 3 with 1 supplement
Out of sample performance of HippUnfold, Automatic Segmentation of Hippocampal Subfields (ASHS), and Freesurfer (FS7).

(A) Side-by-side comparison of results obtained from each method from one representative individual from the Human Connectome Project-Aging (HCP-A) datasets, which was not seen during training. (B) …

Figure 3—figure supplement 1
Additional comparisons of results obtained from Freesurfer (FS7), Automatic Segmentation of Hippocampal Subfields (ASHS), and HippUnfold in 100 Human Connectome Project- Aging (HCP-A) subjects.

All three methods showed a moderate correlation with age, as expected based on previous literature. Volumetric comparison of each method to HippUnfold directly revealed that there is a strong …

Figure 4 with 1 supplement
Comparison of HippUnfold, Automatic Segmentation of Hippocampal Subfields (ASHS), and Freesurfer (FS7) subfield segmentations in native and unfolded space.

Sagittal and coronal slices and 3D models are shown for one representative subject. Note that for HippUnfold hippocampal subfields are the same for all individuals in unfolded space, but for ASHS …

Figure 4—figure supplement 1
Comparison of HippUnfold and fully manual subfield segmentations (data from Thom, 2014) in native and unfolded space from one representative subject.

Sagittal and coronal slices and 3D models are shown for one representative subject. Note that the 3D model of a fully manual segmentation shows clear anterior and posterior digitations which were …

Test set performance in Dice overlaps between HippUnfold and manually unfolded subfields.

All values are compared to ground truth manually defined tissues followed by unfolded subfield definition (manual unfold) to determine how small differences in grey matter parcellation propagate …

Examination of HippUnfold performance on additional datasets Human Connectome Project-Aging (HCP-A) (T1w and T2w) and anisotropic 7T-TSE data.

(A) Sample subjects’ HippUnfold subfield segmentation in native resolution. The first two rows come from the same subjects but using different input data modalities. (B) HippUnfold results from a 7 …

Automated error flagging via overlap with coarse, registration-based segmentation.

(A) Subjects flagged for quality assurance from each dataset based on Dice overlap with a reference mask approximated via deformable registration. (B) Failed subject example illustrating missed …

Tables

Table 1
MRI datasets used in training, evaluation, and comparison to extant methods.

Methods employed include those proposed here (HippUnfold), the same processing but with manual segmentation (similar to previous work; DeKraker et al., 2020) (manual unfold), Freesurfer v7.2.0 (FS7) …

NameModalitiesResolutionSample size (L+R)Methods employed
HCP-YAT1w, T2w0.7 × 0.7 × 0.7 mm3n=590 (training)HippUnfold
Manual unfold
n=148 (testing)HippUnfold
Manual unfold
FS7
HCP-AT1w
T2w SPACE
T2w TSE
0.8 × 0.8 × 0.8 mm3
0.8 × 0.8 × 0.8 mm3
0.4 × 0.4 × 2.0 mm3
n=1312 for T1w, T2w SPACE
n=200 for T2w TSE (FS7, ASHS)
n=200 for T1w (HippUnfold)
HippUnfold
FS7
ASHS
7T-TSE (from ASHS atlas)T2w0.4 × 0.4 × 1.0 mm3n=70HippUnfold
Manual subfields

Additional files

Supplementary file 1

HippUnfold Documentation.

This document fully describes the HippUnfold installation, command-line interface, options, outputs, and provides several useful pieces of information including worked examples and useful tips on viewing data in other common platforms.

https://cdn.elifesciences.org/articles/77945/elife-77945-supp1-v2.pdf
Supplementary file 2

Side-by-side snapshot comparison of Human Connectome Project-Aging (HCP-A) segmentations results from HippUnfold, Freesurfer (FS7), and Automatic Segmentation of Hippocampal Subfields (ASHS) from the left hemisphere.

Snapshots were taken at the conronal centroid, centroid + 15 slices, centroid + 30 slices, and the sagittal centroid.

https://cdn.elifesciences.org/articles/77945/elife-77945-supp2-v2.pdf
Supplementary file 3

Side-by-side snapshot comparison of Human Connectome Project-Aging (HCP-A) segmentations results from HippUnfold, Freesurfer (FS7), and Automatic Segmentation of Hippocampal Subfields (ASHS) from the right hemisphere.

Snapshots were taken at the conronal centroid, centroid + 15 slices, centroid + 30 slices, and the sagittal centroid.

https://cdn.elifesciences.org/articles/77945/elife-77945-supp3-v2.pdf
Supplementary file 4

Detailed mathematical formulation of algorithms used throughout HippUnfold.

https://cdn.elifesciences.org/articles/77945/elife-77945-supp4-v2.pdf
Transparent reporting form
https://cdn.elifesciences.org/articles/77945/elife-77945-transrepform1-v2.docx

Download links