Figure 1.
Step-wise summary of the framework.
Pipeline of the framework is shown at the top of the image. Below the pipeline are representative images of results obtained after each processing step of the framework when applied to an unlabelled image. (A) Brain extraction – create brain mask for bias field correction; (B) Dilate mask to include contrast of brain tissues and CSF for image registration; (C, D) Images before and after bias field correction; (E) Structural parcellation result after single-atlas segmentation propagation; (F) Structural parcellation result after multi-atlas label fusion.
Figure 2.
Parameter optimisation for atlas database with left/right hemisphere separated.
The overall Dice similarity coefficient across all structures resulted from the selection of different number (from 3 to 9) of top-ranked atlases for label fusion. The error bars represent the standard deviation of 12 tests with different Gaussian kernel standard deviation in the LNCC image similarity measurement (from 1 to 6 with 0.5 step increment). The small variation indicates little effect of the Gaussian kernel width towards the overall accuracy.
Figure 3.
Sample images from the cross validation result of the pipeline on the atlas databases.
Parcellation results obtained with the proposed method and parameters. (A) The original MR image from the atlas (B, D) The MR image from the atlas overlaid with corresponding manually labelled anatomical structures which is considered as gold standard. (C, E) The same MR images overlaid with the structural parcellation result after applying our multi-atlas framework. Top row: coronal view, bottom row: axial view.
Figure 4.
Cross validation result on the in vivo mouse brain atlas MRM NeAt [3].
Comparison of the average Dice similarity coefficient using our framework, a single-atlas segmentation propagation method and the STAPLE algorithm. Two-tailed paired t-tests were performed, with multiple comparisons of 40 structures corrected with false discovery rate set to 5%. Error bars representing standard deviation (*: significant difference was discovered between single-atlas method and STEPS algorithm; #: significant difference was discovered between STAPLE and STEPS algorithm).
Figure 5.
The structural parcellation result of applying our multi-atlas framework to a new dataset.
(A) The MR image from an NUS mouse atlas which is treated as a new dataset. (B) MR image of the unlabelled image overlaid with corresponding manually labelled anatomical structures considered as gold standard. (C) The same MR images overlaid with the structural parcellation result after applying our multi-atlas framework. Top row: coronal view, bottom row: axial view.
Figure 6.
Validation on the ability of the multi-atlas framework to parcellate structures of the new dataset.
The new dataset is adopted from the NUS mouse atlas [5] with the corresponding manual labels regarded as gold standard. 12 manually segmented structural labels were included in the comparison which appeared in both of the two atlas databases. Previously obtained optimised parameter combination for the MRM NeAt atlas database were used to calculate the Dice similarity coefficient. Two-tailed paired t-tests were performed, with multiple comparisons of 24 structures corrected with false discovery rate set to 0.05. Error bars representing standard deviation (*: significant difference was discovered between single-atlas method and STEPS algorithm; #: significant difference was discovered between STAPLE and STEPS algorithm).
Figure 7.
Sample images comparing the parcellation result of our framework and the single-atlas based methods.
The selected slices demonstrated that despite some local misalignments in the single-atlas based method (as shown in red arrows). The STEPS label fusion algorithm in our framework successfully preserved the correct local registration in different regions. Structural parcellations are overlaid on the original image (in both coronal and sagittal view, a). (b) Structural parcellation using the proposed framework. (c) Structural parcellation result of a single-atlas based method with part of the cerebellum mis-segmented. (d) Another structural parcellation result of single-atlas based method with the edge between olfactory bulb and cortex mis-segmented.
Figure 8.
Statistical comparison of the structural volume difference between groups of Tc1 Down Syndrome mouse and wild type.
Volumetric comparison on the a) unnormalised data; b) data normalised by total intracranial volume. A two-tailed paired t-test was performed on each of the 40 structures. Multiple comparisons are corrected with false discovery rate q = 0.05. Error bars representing standard deviation. (*: significant difference was discovered between the wild type and the transchromosomic group.).
Table 1.
Statistical significant result on the volumetric comparison between groups of Tc1 Down Syndrome mouse and wild type, result obtained both from our multi-atlas framework as well as the single-atlas based method using all atlases in the database.