Infant FreeSurfer: An automated segmentation and surface extraction pipeline for T1-weighted neuroimaging data of infants 0-2 years

Abstract

The development of automated tools for brain morphometric analysis in infants has lagged significantly behind analogous tools for adults. This gap reflects the greater challenges in this domain due to: 1) a smaller-scaled region of interest, 2) increased motion corruption, 3) regional changes in geometry due to heterochronous growth, and 4) regional variations in contrast properties corresponding to ongoing myelination and other maturation processes. Nevertheless, there is a great need for automated image-processing tools to quantify differences between infant groups and other individuals, because aberrant cortical morphologic measurements (including volume, thickness, surface area, and curvature) have been associated with neuropsychiatric, neurologic, and developmental disorders in children. In this paper we present an automated segmentation and surface extraction pipeline designed to accommodate clinical MRI studies of infant brains in a population 0-2 year-olds. The algorithm relies on a single channel of T1-weighted MR images to achieve automated segmentation of cortical and subcortical brain areas, producing volumes of subcortical structures and surface models of the cerebral cortex. We evaluated the algorithm both qualitatively and quantitatively using manually labeled datasets, relevant comparator software solutions cited in the literature, and expert evaluations. The computational tools and atlases described in this paper will be distributed to the research community as part of the FreeSurfer image analysis package.

Trial registration: ClinicalTrials.gov NCT02058225.

Keywords: Brain surface; FreeSurfer; Infant; MRI; Segmentation.

Copyright © 2020 The Author(s). Published by Elsevier Inc. All rights reserved.

Figures

Fig. 1.

Major image processing steps in the standard FreeSurfer reconall pipeline. Red boxes indicate the ones that are different and were specifically modified in the case of the infant-specific tools.

Fig. 2.

The proposed graphical model for our multi-atlas label fusion tool. Plates indicate replication, shaded variables are observed.

Fig. 3.

Age distribution at scan of the twenty-six subjects in the training data set. Red color indicates data samples that had GM/WM separation drawn by the manual labelers.

Fig. 4.

Skull-stripping results: (left) unprocessed images from BCH_0-2yr data set and (right) intensity normalized and skull-stripped results. Both sets are age sorted, displayed in coronal view and aligned using affine registration to an unbiased spatial coordinate space for easier visualization.

Fig. 5.

Five automated segmentation examples where manual segmentation also contained GM/WM separation: (from top to bottom) newborn, 8mo, 12mo, 16mo, 18mo. From left to right: normalized and skullstripped T1-weighted input image, manual segmentation, manual segmentation outline, automated segmentation, automated segmentation outline. All segmentations (or their outlines) are overlaid on the normalized and skullstripped T1-weighted input image. The segmentation colors correspond to the default Freesurfer colortable.

Fig. 6.

Five automated segmentation examples where manual segmentation did not contain GM/WM separation: (from top to bottom) 2mo, 3mo, 5mo, 6mo, 9mo. From left to right: normalized and skull-stripped T1-weighted input image, manual segmentation, manual segmentation outline, automated segmentation, automated segmentation outline. All segmentations (or their outlines) are overlaid on the normalized and skull-stripped T1-weighted input image. The segmentation colors correspond to the default Freesurfer colortable.

Fig. 7.

Generalized Dice overlap coefficient summary for all subjects and training set sizes (selected by age). The generalized Dice coefficients are displayed for training set sizes 1–25 for all of our subjects, in an age-sorted manner: Subj8 (newborn) → Subj 25 (18 mo).

Fig. 8.

Mean generalized Dice overlap coefficient summary over all training set sizes (selected by age) for all subjects, grouped into five non-overlapping age groups (newborns (N = 5), 2–4 month (N = 4), 5–8 month (N = 5), 9–14 month (N = 6) and 15–18 month olds (N = 6)). The measures are displayed for training set sizes 1–25.

Fig. 9.

Maximum generalized Dice overlap coefficient over all training set sizes (selected by age) for all subjects grouped into five non-overlapping age groups (newborns (N = 5), 2–4 month (N = 4), 5–8 month (N = 5), 9–14 month (N = 6) and 15–18 month olds (N = 6)). The measures are displayed for training set sizes 1–25.

Fig. 10.

Best performing training set sizes (selected by age) computed using generalized Dice coefficients in five non-overlapping age categories (newborns (N = 5), 2–4 month (N = 4), 5–8 month (N = 5), 9–14 month (N = 6) and 15–18 month olds (N = 6)).

Fig. 11.

Generalized Dice score vs age-at-scan computed on the training data set for neighborhood size 5.

Fig. 12.

MANTIS segmentation comparison: (top left) T2w input images (bottom left) MANTIS segmentations, (top right) corresponding T1w input images (bottom right) our segmentation outcome after grouping left/right hemisphere labels together. The list of commonly identified labels are: cerebral cortex, cerebral white matter, deep gray matter, hippocampus, amygdala, cerebellum and brainstem. For more detailed label correspondences see Appendix Table 4.

Fig. 13.

Dice coefficients computed between our segmentations and MANTIS for labels that are commonly identified by these tools: cerebral cortex, cerebral white matter, deep gray matter, hippocampus, amygdala, cerebellum and brainstem.

Fig. 14.

iBEAT subcortical segmentation comparison: (left) T1w input images (middle) iBEAT segmentations, (right) our segmentation outcome. The list of commonly identified labels are: left/right thalamus, caudate, putamen, pallidum, hippocampus and amygdala. For more detailed label correspondences see Appendix Table 5.

Fig. 15.

Dice coefficients computed between our segmentations and iBEAT for labels that are commonly identified by these tools: left/right thalamus, caudate, putamen, pallidum, hippocampus and amygdala.

Fig. 16.

T1w input images of the first forty subjects constituting the recent data elease of the “The Developing Human Connectoe Project” dHCP project viewed in the coronal plane in an unbiased common affine coordinate system.

Fig. 17.

The first forty subjects constituting the recent data release of the “The Developing Human Connectoe Project” dHCP project viewed in the coronal plane in an unbiased common affine coordinate system: (top) skull-stripped input images, and (bottom) segmented images using our new pipeline.

Fig. 18.

The first forty subjects constituting the recent data release of the “The Developing Human Connectoe Project” dHCP project viewed in the coronal plane in an unbiased common affine coordinate system: (top) original T2w input images, and (bottom) dHCP released tissue segmentation outcomes (based on T2w images).

Fig. 19.

Mean Dice overlap measures computed on the DHCP dataset per segmentation labels: (top) “all” segmentation labels and (bottom) “tissue” segmentation labels released by the dHCP consortium.

Fig. 20.

Mean Dice overlap measures computed on the 10 training subjects of the iSEG data set per segmentation labels: (top) the proposed pipeline was run after eliminating the brainstem and cerebellum and segmentation labels combined to match those included in iSEG using neighborhood size 1–12 and (bottom) the volumetric segmentation part of our proposed pipeline was run using the iSEG training data set using neighborhood size 1–9.

Fig. 21.

Surfaces generated for five sample subjects from our training dataset: (from top to bottom) newborn, 8mo, 12mo, 16mo, 18mo. From left to right: left hemisphere white surface, pial surface and spherical representation with a curvature map overlay.

Fig. 22.

Boxplot displays comparing surface measures on the dHCP data set, per hemisphere, all labels combined: (top left) mean absolute distance, (top right) mean sulcal depth difference, (bottom left) mean cortical thickness difference and (bottom right) mean curvature differences.

Fig. 23.

Examples of validation surface points placed on the right and left hemispheres of a randomly selected T1 weighted image from the dHCP data set, viewed on different coronal slices. (Left) validation surface points from both the pial (blue) and white (red) surfaces are shown along with our surface reconstruction solutions (light green – pial surface, yellow – white matter surface); (Right) validation surface points from the white (red) surface are shown along with the dHCP and our white matter surface reconstruction solutions (light blue – dHCP, dark blue – ours).

Fig. 24.

Age-at-scan vs the Tenengrad image sharpness metric (fmeasure) computed on the training dataset (BCH0–2 years).

Fig. 25.

Generalized Dice score for each of training subjects using 4 and 5 as training neighborhood sizes vs the input image volumes’ Tenengrad metric (fmeasure).

Fig. 26.

Generalized Dice score for each of dHCP subject using 5 as training neighborhood sizes vs the input image volumes’ Tenengrad metric (fmeasure): “all” (in red) and common-with-our-pipeline “tissue” labels (in blue).