Tongue Contour Extraction from Ultrasound Images Using Image Parts

Abstract

In this work we propose automatic tongue contour extraction from ultrasound images based on Convolutional Neural Networks (CNN). The developed deep neural network was trained with ultrasound image parts. In training step the distances of the image patch centers from nearest tongue contour was used as the class label. By taking advantage of the fact that images consist of video frames, the time information was used as training input parameter. In the test phase, the image parts obtained by using the sliding window method are given as input to the training model and distance classes are obtained. A class that is close to zero means a position close to the tongue contour. Candidate points for the tongue contour were created using the centers points of the parts. The tongue contour is extracted by fitting a 3rd degree polynomial on center points of candidates. The work is completed by comparing the manually marked contour with the automatically found contour. The dataset consists of ultrasound videos taken in the experimental environment.