Deep-learning-based segmentation of the vocal tract and articulators in real-time magnetic resonance images of speech

Abstract

Highlights A method to segment the vocal tract and articulators in MR images was developed. Median accuracy: Dice coefficient of 0.92; general Hausdorff distance of 5mm. Developed to facilitate quantitative analysis of the vocal tract and articulators. Intended for use in clinical and non-clinical studies of speech. A novel clinically relevant segmentation accuracy metric was also developed.

A segmentation-informed deep learning framework to register dynamic two-dimensional magnetic resonance images of the vocal tract during speech

“…The proposed framework includes a deep-learning-based method to estimate segmentations of the following six anatomical features in the image pair: the head, soft palate, jaw, tongue, vocal tract and tooth space. This method is described in [56] and consists of two steps. First, segmentations of the six anatomical features in the image pair are estimated using a pre-trained CNN.…”

“…Second, a connected-component-based post-processing step is performed to remove anatomically impossible regions from the segmentations. For full information about the segmentation method, the reader is referred to [56] .…”

“…Several methods to segment articulators in dynamic 2D MR images of the vocal tract during speech have been developed [54] , [55] , [56] , [57] , [58] , [59] , [60] , [61] , [62] , [63] , [64] . However, only one of these fully segments several groups of articulators in the images [56] .…”

“…A metric based on velopharyngeal closure has been proposed and used to evaluate the accuracy of a method to segment dynamic 2D MR images of the vocal tract during speech [56] . This metric quantifies how many of the velopharyngeal closures in the ground-truth (GT) segmentations occur in the estimated segmentations, and is calculated by comparing corresponding consecutive segmentations in the two series.…”

An open-source toolbox for measuring vocal tract shape from real-time magnetic resonance images

“…Furthermore, those approaches which are available have not yet been demonstrated to generalise beyond the individual datasets for which they were developed. These methodologies typically involve automated or machine learning processes which are trained and tested against a narrow range of data, typically composed of a small number of speakers scanned at a single imaging centre [12][13][14][15][16][17][18]. The development of these techniques has sampled disproportionally from a single image repository [6].…”