Fig 1.
To create a template, an input image (template subject) undergoes a sequence of processing steps (black boxes) such as pre-processing, template construction, labelling and registration to finally output a labelled image. Red boxes represent input and output images.
Table 1.
Target Areas.
Table 2.
Segmentation Metrics.
Fig 2.
Atlas—Template and Associated Labels.
Segmented Atlas showing (1) Ventricles, (2) Motor Cortices, (3) Hippocampi, (4) Thalami, (5) Caudate (6) Amygdala and (7) Cerebellum illustrating the (a) transverse, (b) parasagittal, (d) dorsal planes and (c) 3d rendering of labels on a ventrolateral view with rostral to the left of the image.
Fig 3.
The (a) skulled template and (b) deskulling mask.
Variability displayed in extracortical areas reinforces the reason why deskulling is an important step in pre-processing.
Table 3.
Deskulling Metrics.
Fig 4.
Dice Similarity Coefficients of all structures in the low resolution template.
Values range from 0.12 for the amygdala to 0.75 for the cerebellum. Red = Ventricles, Green = Right Motor Cortex, Dark Blue = Left Motor Cortex, Yellow = Hippocampus, Light Blue = Thalamus, Purple = Caudate Nucleus, Peach = Amygdala, Grey = Cerebellum, Brown = Intraobserver Error. Error bars show the standard deviation of the mean for all compared structures. Intraobserver Error was calculated by comparing manual segmentations of all seven labelled structures in the four test subjects. Number of test subjects (n = 4).
Fig 5.
Box plot displaying the average of all comparisons performed across all segmented structures.
Shows the median Dice Similarity Coefficient to be 0.61 with the False Positive and False Negative values having a very large variation. This indicates a wide spread in the accuracy of segmentation with regards to various structures. The Jaccard Coefficient is as expected lower than the Dice proportionally due to single use of the intersect in its calculation leading to a more precise comparison.