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

摘要: Abstract Background and Objective Magnetic resonance (MR) imaging is increasingly used in studies of speech as it enables non-invasive visualisation the vocal tract articulators, thus providing information about their shape, size, motion position. Extraction this for quantitative analysis achieved using segmentation. Methods have been developed to segment tract, however, none these also fully any articulators. The objective work was develop a method multiple groups articulators well two-dimensional MR images speech, overcoming limitations existing methods. Five image sets (392 total), each different healthy adult volunteer, were work. A convolutional network with an architecture similar original U-Net following six regions sets: head, soft palate, jaw, tongue, tooth space. five-fold cross-validation performed investigate segmentation accuracy generalisability network. assessed standard overlap-based metrics (Dice coefficient general Hausdorff distance) novel clinically relevant metric based on velopharyngeal closure. Results segmentations created by had median Dice 0.92 distance 5mm. segmented head most accurately (median 0.99), palate space least coefficients 0.93 respectively). correctly showed 90% (27 out 30) closures sets. Conclusions An automatic successfully developed. intended use clinical non-clinical which involve position In addition, assessing articulator methods