Automatic intracranial space segmentation for computed tomography brain images

Abstract

Craniofacial disorders are routinely diagnosed using computed tomography imaging. Corrective surgery is often performed early in life to restore the skull to a more normal shape. In order to quantitatively assess the shape change due to surgery, we present an automated method for intracranial space segmentation. The method utilizes a two-stage approach which firstly initializes the segmentation with a cascade of mathematical morphology operations. This segmentation is then refined with a level-set-based approach that ensures that low-contrast boundaries, where bone is absent, are completed smoothly. We demonstrate this method on a dataset of 43 images and show that the method produces consistent and accurate results.

Fig. 1

Orthogonal views of an example CT image (a) with the intracranial space boundary marked with white contours. b The midsagittal slice was taken from the same CT image with the frontal fontanelle and the basion and opisthion, which are used to demarcate the foramen magnum, annotated

Fig. 2

Flowchart of mathematical morphology preprocessing steps in order to transform the input CT image into the three output segmentation images: S_bone, S_init, and S_brainstem

Fig. 3

Illustration of threshold selection technique on a histogram of typical image in our dataset (i). The air/background and high-intensity tissue peaks are annotated with asterisks, along with the S_air/S_tissue,low threshold. (ii) A subset of the original histogram and its second derivative. The first subthreshold second derivative value, after the first peak in the second derivative, on the negative and positive sides, is used to choose the S_{tissue, low}/S_{tissue, high} and the S_{tissue, low}/S_bone thresholds (denoted by asterisks), respectively

Fig. 4

Orthogonal views of preprocessing steps: (a) original CT image; (b) image after intensity-based segmentation into four classes: air (black), low-intensity tissue (dark gray), high-intensity tissue (light gray), and bone (white); (c) S_opened: the high-intensity tissue segmentation image, S_{tissue, high} after erosion, retention of the largest connected component and dilation; (d) all three output segmentations: S_init (dark gray), S_brainstem (light gray), S_bone (white)

Fig. 5

a–b Results of applying the illusory surface method to three selected images from the dataset. The top row shows a midsagittal slice with the initial segmentations, S_init, denoted by black contours and the final ICS segmentation estimations computed by the illusory surface method shown as white contours. The curved lines highlight the extent of the frontal fontanelles and show that the segmentation boundaries have been slightly contracted and smoothed. The inferiorly located arrowheads in a–c point to locations where the segmentation boundary in the foramen magnum has been smoothed to form a consistent segmentation boundary shape between images. The bottom row shows surface renderings of the final segmentations

Fig. 6

a A midsagittal slice of one image in the dataset showing the effect of extreme values of β = 2 (black) and β = 20 (inner white contour), with the ground truth segmentation (outer white contour). The oversmoothing effect is particular evident in highly curved skull sections, such as the orbits which are annotated by a white arrow. b A midsagittal slice of another image of the dataset showing the typical segmentation error where the estimated boundary (white) of the foramen magnum is inferior to the ground truth boundary (black) (see arrowhead). The parameter choices for this figure were (β = 2, μ = 30). Note that the skulls displayed in this figure are physically distorted by the craniosynostosis; the aspect ratio of these panels is correct

Fig. 7

A midsagittal slice of the image that consistently produced the highest error value across parameter combinations in original orientation (a) and after manual rotation so that the basion/opisthion axis is in the axial plane (b). The estimated segmentation for (β = 2, μ = 30) (white) and ground truth boundary (black) are shown. The largest discrepancy between the ground truth and estimated segmentation boundary in a is indicated by the arrow