EdgeTrak Automatic Contour Tracking in Ultrasound Images

EdgeTrak System

When measuring tongue shapes from  ultrasound images, automatic extraction and tracking of the tongue surface is necessary to avoid time consuming manual extractions. We developed a  system, EdgeTrak, which can track the tongue surfaces through a sequence of 2D ultrasound images. EdgeTrak is currently being used by scientists at several institutions. Feedback from them, and validation, indicate that the system is efficient and accurate for speech research and related applications.

Typical Results

edgetrak_results


Results from a speech sequence

Input

Any sequence of black and white images can be measured using Edgetrak.  Open-contours are extracted from the lower edge of a white to black gradient image.

Potential Input Problems

  • Ultrasound images are quite noisy.
  • There are unrelated high contrast edges in the images.
  • The tongue surface might be interrupted in places.
Formulation

Snakes

We use a novel deformable contour model based on active contour models (snakes). The internal energy is calculated using the contour smoothness The external energy is calculated using the negative of the image gradient and region based intensity difference.

 

edgetrak_smooth_small 

 
 Calculation of smoothness
Novelty of the EdgeTrak System

In a snake model, the external energy is usually related to the gradient of the image. In reality, however, images are noisy and there are always unrelated high-contrast edges which make the gradient information  insufficient to extract edges of interest.  In an ultrasound image of the human tongue, a bright white band is formed by the reflection of ultrasound waves at the upper surface of the tongue. The lower edge of the band is the upper surface of the tongue, and the upper edge of the band has no physical interpretation. Thus, only the lower edge is of interest to speech scientists though both edges have high gradient. It is hard to distinguish them by only using  gradient information .

edgetrak_tongue_examples
An example of  ultrasound image of the tongue

To distinguish the tongue surface from the other high contrast edges, e.g., the upper edge of the air reflection, we introduce a constraint which is related to intensity in local regions. 

edgetrack_region_of_definition
Local region definition

Please visit the EdgeTrak web page at the VIMS Lab, University of  Delaware.
Send any questions to mli@cis.udel.edu