Segmentation of tongue muscles from super-resolution magnetic resonance images

Abstract

Imaging and quantification of tongue anatomy is helpful in surgical planning, post-operative rehabilitation of tongue cancer patients, and studying of how humans adapt and learn new strategies for breathing, swallowing and speaking to compensate for changes in function caused by disease, medical interventions or aging. In vivo acquisition of high-resolution three-dimensional (3D) magnetic resonance (MR) images with clearly visible tongue muscles is currently not feasible because of breathing and involuntary swallowing motions that occur over lengthy imaging times. However, recent advances in image reconstruction now allow the generation of super-resolution 3D MR images from sets of orthogonal images, acquired at a high in-plane resolution and combined using super-resolution techniques. This paper presents, to the best of our knowledge, the first attempt towards automatic tongue muscle segmentation from MR images. We devised a database of ten super-resolution 3D MR images, in which the genioglossus and inferior longitudinalis tongue muscles were manually segmented and annotated with landmarks. We demonstrate the feasibility of segmenting the muscles of interest automatically by applying the landmark-based game-theoretic framework (GTF), where a landmark detector based on Haar-like features and an optimal assignment-based shape representation were integrated. The obtained segmentation results were validated against an independent manual segmentation performed by a second observer, as well as against B-splines and demons atlasing approaches. The segmentation performance resulted in mean Dice coefficients of 85.3%, 81.8%, 78.8% and 75.8% for the second observer, GTF, B-splines atlasing and demons atlasing, respectively. The obtained level of segmentation accuracy indicates that computerized tongue muscle segmentation may be used in surgical planning and treatment outcome analysis of tongue cancer patients, and in studies of normal subjects and subjects with speech and swallowing problems.

Keywords: Atlasing; Game theory; Human tongue; Magnetic resonance imaging; Segmentation.

Fig. 1

An example of a super-resolution 3D MR image of the tongue, reconstructed from sets of orthogonal (a) sagittal, (b) coronal and (c) axial MR images with a limited field of view. The unshaded areas correspond to individual images, the lightly shaded areas to the intersection of two orthogonal images, and the strongly shaded areas to the intersection of three orthogonal images. (d) As a result, only the tongue region is intersected by all orthogonal images, whereas the corners of the volume are not covered by any orthogonal image and therefore represent blank areas.

Fig. 2

A schematic illustration of the game-theoretic framework for landmark-based segmentation of tongue muscles from 3D MR images.

Fig. 3

(a) Nine different types of Haar-like features are used to generate appearance likelihood maps. The feature response is the difference between voxel intensities inside shaded and unshaded regions. (b) Haar-like features are computed at 125 voxels (shaded) of the 13³-voxels large landmark neighborhood.

Fig. 4

Atlasing results for a selected super-resolution 3D MR image of the tongue, shown in a sagittal cross-section and obtained by applying (a) the game-theoretic framework, (b) B-splines atlasing and (c) demons atlasing. In the first row, the results are shown as semi-transparent domains, where the red color indicates the accumulated propagations of reference segmentations from images in the training set, while the green color indicates the majority voting of the accumulated propagations. In the second row, the results are shown as colored volumes, where the yellow color indicates the segmented genioglossus muscle, while the blue color indicates the segmented inferior longitudinalis muscle

Fig. 5

Segmentation of genioglossus and inferior longitudinalis tongue muscles for a selected super-resolution 3D MR image, shown in sagittal (top), coronal (middle) and axial (bottom) cross-sections, obtained by (a) manual segmentation, (b) game-theoretic framework, (c) B-splines atlasing and (d) demons atlasing (d). The brown color indicates either the manual segmentation (a) or the overlap between manual and computerized segmentations (b–d) of the genioglossus muscle. The green color indicates either the manual segmentation (a) or the overlap between manual and computerized segmentations (b–d) of the inferior longitudinalis muscle. The red color indicates the disagreement between manual and computerized segmentations.