Tracing the tongue with GLoSsatron

1
Tracing the tongue with GLoSsatron

• Adam Baker, Jeff Mielke, Diana Archangeli
• University of Arizona
• Supported by
• College of Social and Behavioral Sciences,
  University of Arizona
• James S. McDonnell Foundation 220020045 BBMB

2
The Need

• Taking point measurements from ultrasound images
  is tedious and time-consuming.
• even when simple methods are used
• easily 75 of the time required to run an
  experiment
• Obtaining measurements automatically would
  ameliorate that problem.

3
The Problem

• There are a number of features that make
  ultrasound images difficult to measure
  automatically.
• A tour of the problem

4
Rarely this nice
5
Potentially Ill-formed Lines
?
6
Graininess
7
Beamforming artifacts
8
Variable illumination
9
Phantom palates
Really an ultrasound artifact
10
Technology vs. Biology

• Problems are attributable to
• ultrasound technology
• speaker idiosyncrasies
• hydration level that day
• muscle morphology
• pressure applied to transducer
• waddle (good)
• scruff (bad)

11
Technology vs. Biology

• The magnitude of the problem can be reduced
  considerably if we have high standards for our
  subjects.
• This is a more practical solution for studies of
  English speakers than for work in other
  languages.
• I suggest that a goal of automatic edge detection
  should be an algorithm that works (fairly well)
  for non-ideal images.

12
GLoSsatron

• GLoSsatron is a system intended to produce
  quality surfaces
• for a wide range of image qualities
• with a minimum of input from the experimenter

13
GLoSsatron

• It is named for the three filters used to enhance
  the tongue surface.
• Gaussian
• Laplacian
• Sobel
• Why are filters needed at all?

14
Too many edges

• Sobel filter finds the gradient of the image
• i.e. parts where theres a change from light to
  dark
• Almost useless in such a high noise situation

15
1. Reducing noise

• A Gaussian convolution is used to eliminate
  noise.
• Every pixel is
• replaced by
• a weighted sum
• of itself and its
• neighbors.

16
2. Reducing noise

• The tongue
• surface becomes
• more prominent with respect to the noise in the
  image.
• This is equivalent to a low-pass filter.

17
2. Enhancing the Edge

• A Laplacian filter is used to enhance the
• remaining edges
• The process
• of convolution
• is identical.
• This is the 2nd
• derivative of the
• Gaussian.

18
2. Enhancing the Edge

• The tongue surface is now quite prominent w.r.t
  the rest of the image.
• The task now is to identify the tongue surface.

19
3. Zeroing In

• At this point the Sobel (gradient) filter
  becomes helpful.
• The tongue surface is now quite prominent.

20
Searching for the surface

• To find the surface we use a radial grid, we
  search along predefined radii.

21
Searching Along a Radius

• Search in a user-defined portion of the radius.

22
Searching Along a Radius

• Find the maximum point of the Laplacian

23
Searching Along a Radius

• Find the corresponding point on the Sobel.

24
Searching Along a Radius

• Find the first lower maximum on the Sobel.

25
Searching Along a Radius

• This is the point we want.

26
Searching for the surface

• This heuristic is quite simple.
• A more sophisticated technique will almost
  certainly yield superior results.
• However, much is to be gained in post-processing.

27
Catching Errors

• No edge detection system will score 100

Small Gaps
No tongue to find
28
Catching Errors

• This algorithm misses three real points, and
  falsely identifies many non-tongue points.

29
Catching Errors

• These are outliers relative to their neighbors
  this can be quantified.

30
Catching Errors

• They can be detected and eliminated, either with
  simple or complex means.

31
Catching Errors

• Experience so far eliminating false data points
  is the easiest and most rewarding way to increase
  the edge detection accuracy.
• So how about those bad images?

32
Rarely this nice
33
Rarely this nice
34
Potentially Ill-formed Lines
35
Potentially Ill-formed Lines
?
36
Potentially Ill-formed Lines
37
Graininess
38
Graininess
39
Beamforming artifacts
40
Beamforming artifacts
41
Variable illumination
42
Variable illumination
43
Phantom palates
Really an ultrasound artifact
44
Phantom palates
45
Conclusion

• GLoSsatron is a new algorithm that can be
  efficiently implemented for users.
• The experimenter will supply only a
  subject-specific search window (i.e. where the
  tongue is going to appear).
• This program, as with others like it, has the
  potential to save experimenters tremendous
  quantities of time.