Figure 1.
Depiction of two charts and the associated diffeomorphisms chosen to illustrate the interpretation.
The charts are related via diffeomorphic coordinate transformations as depicted in the figure, in which points X, Y in the hippocampus chart and the amygdala chart are compared using the forward and inverse mappings. In our paper the charts are manually delineated structures including the amygdala, caudate, hippocampus, putamen, thalamus, lateral ventricle, the 3rd ventricle, and the 4th ventricle.
Table 1.
Three different groups of MR scans and their respective size, age range, resolution, imaging protocol, and pathology.
Figure 2.
A comparison of two sets of atlas segmentations.
Panels A–C show the mean Dice values and its standard deviations of the sixteen structures, for the three different groups, obtained from likelihood-fusion via multi-atlas LDDMM based on two different sets of atlas label definitions: 1) sixteen manually defined structures (red); 2) sixteen manually defined structures and three generic neighboring tissue segmentations (green).
Figure 3.
A comparison of segmentation accuracy between single-atlas LDDMM and multi-atlas LDDMM.
Panels A, B, C respectively show the mean Dice overlaps and the standard deviations of the sixteen structures obtained from single-atlas LDDMM (red) and likelihood-fusion via multi-atlas LDDMM (green) for the three different groups.
Figure 4.
Depiction of the variability within different single atlases.
Scatterplot of Dice overlaps of automated segmentations of sixteen different structures of one subject from 6 different atlases using single atlas LDDMM.
Figure 5.
Example of subcortical segmentations from single- and multi-atlas LDDMM approaches.
Panel A shows the automated segmentation results of two subjects using single-atlas LDDMM, while panel B shows the segmentation results for the same subjects using multi-atlas LDDMM approach.
Figure 6.
Boxplots of the Dice overlaps for sixteen different structures.
The Dice overlaps were computed between the automated segmentations of sixteen different structures of one subject from 28 different atlases using single atlas LDDMM and the one from multi-atlas LDDMM (blue dotted line).
Figure 7.
The convex weighting function normalized over each structure.
For each structure, we color-coded the quantity that is depicted for each atlas (column) and each of the 16 structures (rows).
Figure 8.
The likelihood contribution of each atlas averaged over the atlas structures.
The quantity was obtained from the 28 different atlases (columns) for each structure (row).
Figure 9.
Example slices for a comparison of multi-atlas LDDMM, STAPLE, and Spatial STAPLE.
Three representative 2-D slices of three structures near medial temporal regions – the amygdala, the hippocampus, and the ventricle in both hemispheres obtained respectively from manual delineation (top row), likelihood-fusion via multi-atlas LDDMM (2nd row), STAPLE (3rd row), and Spatial STAPLE (bottom row).
Table 2.
The average Dice overlaps between manual volume and the automated volume measured over the fourteen datasets of the first group for each structure for comparisons of STAPLE, Spatial STAPLE, and likelihood-fusion via multi-atlas LDDMM.
Table 3.
Mean and standard deviations of Dice overlaps computed across the fifteen subjects in the second group for each structure for comparisons of STAPLE, Spatial STAPLE, and likelihood-fusion via multi-atlas LDDMM.
Table 4.
Mean and standard deviations of Dice overlaps obtained respectively from STAPLE, Spatial STAPLE, and likelihood-fusion via multi-atlas LDDMM in segmenting the 16 structures of the six subjects in the third group.