Fig 1.
Illustration of 3 ontology levels of the segmentation results from our multi-atlas brain segmentation pipeline within the same subject.
The structures defined by red name are indicated in the images in red arrows. The name of each structure is listed in Table 1.
Fig 2.
The schematic workflow of the segmentation pipeline.
Given 5 atlases in use, 3 atlases were selected through the pre-selection step to estimate the final segmentation labels of the target image.
Fig 3.
The coarse 7-label segmentation for a single subject.
The 7 labels include the CSF space with red boundary, the gray matter with blue boundary, the white matter with yellow boundary, the ventricles with pink boundary, tissues in the skull base (optic chiasm, optic nerves, pituitary gland, and other tissues which have similar intensities as the adjacent brain tissues) with cyan boundary, skin (includes lipid layer) with green boundary, and the remaining space with orange boundary.
Table 1.
ROIs for evaluation test under Level A, Level B and Level C.
Fig 4.
Dice overlap of Level C brain structures for all population groups with the 4L, LV, MI and random pre-selection method.
Fig 5.
Dice overlap of Level A, Level B and Level C structures for Top 10 with the 4L, LV, MI and random pre-selection method.
Fig 6.
The ratio of Dice overlap < 0.7 in 286 structures for all subjects with the 4L, LV, MI and random pre-selection method.
Fig 7.
Examples of 5 selected atlases for a geriatric subject with different atlas pre-selection approaches.
The same atlases shared across different pre-selection approaches are indicated with the same border color.
Fig 8.
Examples of 5 selected atlases for an adult subject with different atlas pre-selection approaches.
The same atlases shared across different pre-selection approaches are indicated with the same border color.