Segmentation and 3D reconstruction of the vocal tract from MR images – a comparative study

Speech production is an important human function involving a set of organs with specific morphological and dynamic aspects. The inter-speaker variability, the coarticulation or the nasality are some interesting aspects to improve a realistic 3D modeling of the vocal tract. For this, the understanding of the mechanism of speech production is crucial, as the current image data is not sufficient to reproduce truthfully the speaker’s anatomy and articulation. Hence, the goal of 3D modeling is to generate the complete geometrical and dynamical information concerning the vocal tract from medical images, such as from magnetic resonance imaging (MRI). This work aims to describe and compare two different segmentation techniques to attain the 3D shape of the vocal tract during speech production from MR images: the former based on manual tracing of the vocal tract contours and the latter based on image thresholding. Thus, the segmented cross-sectional areas were measured, and 3D models were built from the sagittal data by blending the contours obtained from the two segmentation techniques. The mean error of the measures computed were low for both segmentation techniques, which let us conclude that the techniques are useful to evaluate the vocal tract geometry accurately. Additionally, the 3D models built using both segmentation techniques were also very similar and truthful. However, when the coronal data was used, various difficulties occurred. Despite the level of automation in performing the image processing tasks, the analysis of vocal tract from images remains complex, given the various problems that persist. Some of these problems are related with the MR image acquisition technique and with the morphology of the vocal tract (i.e. the nonidentification of teeth and the similarity of signal intensity between the vocal tract and of some of the surrounding structures). According to (Soquet et al. 1998), that presents a comparative study to assess the accuracy of three different segmentation techniques, the image thresholding method revealed an inferior dispersion, but the overall results presented small average error and large error distribution. In this work, we describe and compare two different segmentation techniques to extract the contours to be used in 3D reconstruction of the vocal tract during speech production from MR images: one based on the manual tracing of the segmentation contours, and another based on image threshold. The remaining of this paper is organized as follows. In the next section, the description of the MRI protocol, the speech corpus and the image analysis and assessment are described. Then, in section three, the area measurements obtained by using the two segmentation techniques under comparison are presented and discussed, and the 3D models built for the vocal tract are shown and examined. Finally, the conclusions are pointed out in the last section.