Automatic segmentation of vocal tract MR images

Unsupervised segmentation of the vocal tract from real-time MRI sequences

“…The obtained frame rate of 14 frames/s, for this dataset, although not as high as reported by other authors (e.g., Lammert et al, 2010;Raeesy et al, 2013), has already been useful to support articulatory studies (e.g., Oliveira et al, 2012;Martins et al, 2012) and research on velar movement detection using surface electromyography (Freitas et al, 2014). Therefore, it constitutes an adequate dataset to demonstrate the proposed segmentation method.…”

“…The main goal was to obtain a model that could provide data on the different/relevant shape properties of the vocal tract during speech production to distinguish between different types of articulation. Raeesy et al (2013) perform vocal tract segmentation from RT-MRI images by using oriented active shape models on each image frame separately. They also propose an automatic method to set the landmarks used for training.…”

“…This results in a lack of landmarks in important points, such as the upper lip, leading to poor segmentation results as depicted in Fig. 4 of Raeesy et al (2013). Furthermore, careful positioning of landmarks might be very useful for easier contour partitioning (e.g., identifying different regions of interest, such as the tongue or the velum).…”

“…Evaluation is often qualitative (Bresch and Narayanan, 2009), performed over a small image set (Vasconcelos et al, 2011;Raeesy et al, 2013) or considering annotations by a single observer .…”

“…Considering other ASM/AAM based vocal tract segmentation methods (Vasconcelos et al, 2011;Raeesy et al, 2013) our approach adds to them by: (1) requiring a single manual initialisation step per speaker (as opposed to one initialisation per image); (2) using two models to explicitly tackle oral and nasal configurations; (3) addressing data from multiple speakers (only one speaker considered in Vasconcelos et al, 2011); (4) using criterial positioning of landmarks allowing subsequent identification of different regions of interest in the vocal tract; and (5) starting the segmentation of each frame with the segmentation of the previous frame.…”