See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/234943982

An ultrasonographic periodontal probe

Article · February 2010

DOI: 10.1063/1.3362255

CITATION READS

1 1,125

2 authors, including:

Crystal Bertoncini Acosta

Capital One
15 PUBLICATIONS   192 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Crystal Bertoncini Acosta on 16 May 2014.

The user has requested enhancement of the downloaded file.

AN ULTRASONOGRAPHIC PERIODONTAL PROBE

C. A. Bertoncini and M. K. Hinders

NDE Lab, College of William and Mary, Williamsburg, VA, 23187-8795

ABSTRACT. Periodontal disease, commonly known as gum disease, affects millions of people. The
current method of detecting periodontal pocket depth is painful, invasive, and inaccurate. As an
alternative to manual probing, an ultrasonographic periodontal probe is being developed to use
ultrasound echo waveforms to measure periodontal pocket depth, which is the main measure of
periodontal disease. Wavelet transforms and pattern classification techniques are implemented in
artificial intelligence routines that can automatically detect pocket depth. The main pattern
classification technique used here, called a binary classification algorithm, compares test objects with
only two possible pocket depth measurements at a time and relies on dimensionality reduction for the
final determination. This method correctly identifies up to 90% of the ultrasonographic probe
measurements within the manual probe's tolerance.

Keywords: Ultrasonography, Signal Analysis, Wavelet Transform, Acoustics, Pattern Classification

PACS: 43.60.Lq, 87.63.D

INTRODUCTION

In the clinical practice of dentistry, radiography is routinely used to detect

structural defects like cavities. However, ionizing radiation is harmful to the patient and
has been shown to eventually lead to cavities via the demineralization of teeth.
Radiography can also only detect defects parallel to the projection path, which makes
cracks are difficult to detect, and it is useless for identifying conditions such as early stage
gum disease because soft tissues are transparent to x-rays. Medical ultrasound, however,
is safe to use as often as indicated, and computer interpretation software can make disease
detection automatic. The structure of soft tissues can be effectively analyzed with
ultrasound, and even symptoms like inflammation can be registered with this technology
[1].

FIGURE 1. The diagram compares a) the manual and b) the ultrasonographic periodontal probes.

CREDIT LINE (BELOW) TO BE INSERTED ON THE FIRST PAGE OF EACH PAPER

EXCEPT FOR ARTICLES ON PP. 295-302, 427-434, 910-917, 1021-1028, 1233-1240, 1452-1459,
1887-1894, and 2015-2022.

CP1211, Review of Quantitative Nondestructive Evaluation Vol.29,

edited by D. O. Thompson and D. E. Chimenti
© 2010 American Institute of Physics 978-0-7354-0748-0/10/$30.00

1566
FIGURE 2. The experimental equipment used to run the periodontal probe are shown here, including a)
laptop computer, b) interface device, which includes water pressure control c) pulser-receiver, d)
ultrasonographic probe, e) manual probe, and f) foot pedal to turn the water on and off.

One application of ultrasound in dentistry, a periodontal probe, is being developed

as a spin-off of NASA technology [2]-[8]. Periodontal disease is caused by bacterial
infections in plaque, and in advanced stages can cause tooth loss when the periodontal
ligament holding the tooth in place erodes [9]. Periodontal disease is so widespread
worldwide that 10-15% of adults have advanced stages of the disease with deep
periodontal pockets that put them at risk of tooth loss [10]. The usual method of detection
is with a thin metal probe scribed with gradations marking depth in millimeters (Figure
1a). The dental hygienist inserts the probe into the area between the tooth and gum to
estimate the depth to the periodontal ligament. At best, this method is accurate to +/-1mm
and depends upon the force the hygienist uses to push the probe into the periodontal
pocket. Furthermore, this method is painful and often causes bleeding [11]. The
ultrasonographic periodontal probe uses high-frequency ultrasound to determine the depth
of the periodontal ligament non-invasively. An ultrasonic transducer projects high
frequency (10-15 MHz) ultrasonic energy between the tooth and the gum and detects
echoes of the returning wave (Figure 1b).
In the usual practice of ultrasonography1, the time delay of the reflection is
converted to a distance measurement by using the speed of sound in water (1482 m/s).
However, both experimental and simulation waveforms show that the echoes from the
anatomy of interest are smaller than the surrounding reflections and noise. Previous work
with the ultrasonographic periodontal probe by Hou developed the Dynamic Wavelet
Fingerprint (DWFP) method to transform wavelet coefficients to 2D binary images [8].
The method is general, but when directly applied to data obtained from a 4th generation
ultrasonographic periodontal probe tested on 14 patients [12]-[14], the authors used image
recognition techniques to resolve at best 60% of the pocket depths accurately within a
tolerance of 1mm. This paper describes the development of more sophisticated artificial
intelligence routines that include pattern classification as well as image recognition
techniques to detect pocket depth for the ultrasonographic periodontal probe.

METHOD

Previous publications describe the fabrication of several generations of prototype

ultrasonographic probes [1]-[8], [12]-[14]. The body of the 5th generation probe is
manufactured similarly to other dental hand pieces, with the 10 MHz piezoelectric

1
We use “sonographic” or “ultrasonographic” to imply the diagnostic application of ultrasound energy instead of
“ultrasonic,” which in the dental industry often refers to kHz-frequency vibrations used for cleaning.

1567
FIGURE 3. The distribution of periodontal pocket depths measured with the manual probe from the data set
collected at ODU in May, 2007.

transducer located in the head of the probe. The fabrication of the probe allows water,
which is the ultrasonic coupling agent, to be funneled through the custom-shaped tip. The
rest of the components used to control the probe, including the general purpose pulser-
receiver and the custom built water flow interface device, are shown in Figure 2.
In April and May of 2007, clinical tests were performed at Old Dominion
University's (ODU) Dental Hygeiene Research Center on 12 patients using both the
ultrasonographic periodontal probe and the traditional manual periodontal probe. The
human subjects protocol was approved by Institutional Review Boards at both William and
Mary and ODU. Most of the measurements reflected healthy subjects, with 76% of the
manually-measured data measuring 3mm or less (Fig. 3). A comparison of a waveform
from the ultrasonographic periodontal probe with a waveform generated by 3D parallel
elastodynamic finite integration simulations on the relevant anatomy [15] is shown in
Figure 4. Note that there is no visible reflection from the periodontal ligament in either the
experimental or simulated waveforms.
To detect pocket depth, we will use these basic steps of the artificial intelligence:
1. Feature Extraction: Acquire wavelet fingerprints with DWFP and find fingerprint
properties using image recognition software.
2. Feature Selection: Find the mean and standard deviation of each property for all
waveforms and collect key values where the average property varies per pocket
depth.
3. Binary Classification: Compare the selected features in a leave-one-out routine
using well-known pattern classification schemes and only two possible pocket
depths at a time.
4. Dimensionality Reduction: Evaluate binary labels using four different methods to
collapse each binary choice to one label.
5. Classifier Combination: Combine the predicted labels from the most precise tests to
improve accuracy and/or spread of labels.
Each step is explained in briefly in the sections that follow. It is also important to
note that because of the computation time of this task, the computer algorithms were
adapted to run on William and Mary's Scientific Computing Cluster
(http://www.compsci.wm.edu/SciClone/).

Feature Extraction

Since reflections from the periodontal ligament are not apparent in the ultrasound
waveforms, advanced mathematics are needed to identify pocket depth. The clinical trials
yielded ultrasonic waveforms ws,k(t), where there are k=1,…,30 waveforms recorded for
each tooth site s=1,…,1470. To extract features from ws,k(t) for classification, we use the
Dynamic Wavelet Fingerprinting Technique (DWFP) (Fig. 5), which creates binary

1568
FIGURE 4. A sample waveform from a) simulation and b) experiment are compared here. The large reflections
indicated in the boxes are artifacts from the tip geometry. Logically, the echo from the bottom of the pocket
would occur in between the boxes, but no sizable reflection is visible.

contour plots of the wavelet transform coefficients C(a, b). The DWFP, along with a
mother wavelet ψ(t) and some parameters including translation a (proportional to time)
and scaling b (proportional to frequency) applied to the waveforms ws,k(t) yields an image,
I. Preliminary tests showed that mother wavelets Debauchies 3 (db3) and Symelet 5
(sym5) showed promise for this application, and so both were applied in this technique.
The resulting image I contains fingerprint-like binary contours of the initial waveform
ws,k(t) at tooth site s.
The next step is to perform image processing in MATLAB to gather properties of
the fingerprints in the waveform ws,k(t). First, the binary image I is labeled with 8-
connected objects, allowing each individual fingerprint in I to be recognized as a separate
object [18]. Next, several properties are measured from each fingerprint. Some of these
properties include counting the on- and off-pixels in the region, but many involve finding
an ellipse matching the second moments of the fingerprint and measuring properties of that
ellipse, such as eccentricity. In addition to the orientation measure provided by the ellipse,
another measurement of inclination relative to the horizontal axis was determined by
Horn's method for a continuous 2D object [19]. Further properties were measured by
determining the boundary of the fingerprint and fitting 2nd or 4th order polynomials to
represent the shape of the fingerprint boundary.
The image processing routines result in fingerprint properties ps.k.n[t], where n
represents the image processing-extracted fingerprint properties. These properties are
discrete in time because the values of the properties are matched to the time value of the
fingerprint's center of mass. Linear interpolation yields a smoothed array of property
values, ps.k.n(t).

Feature Selection

It is now possible to begin reducing the dimensionality of the features extracted

from the DWFP technique. One dimension can be eliminated immediately by averaging
over the repeated waveforms, k, so that p=ps,n(t). The remaining dimensionality reduction
will be performed by selecting values of p at times ti for particular fingerprint properties ni
so the feature vector at tooth site s, fs, will be formed of fs(i)=ps,ni(ti). Then the
classification matrix will have rows corresponding to s and columns corresponding to fs(i).

1569
FIGURE 5. The DWFP technique [8] begins with a) the ultrasonic signal, where it generates b) wavelet
coefficients indexed by time and scale. Then c) the coefficients are sliced and d) projected onto the time-
scale plane, forming a binary contour image.

It is also possible to reserve this dimension for a later majority voting method, which will
not be described in detail here.
To select features of interest, we look for values of the fingerprint properties that
are different, on average, for different measured values of pocket depth. Therefore, for
each property n, we find the mean and standard deviation over that property for all tooth
sites:
N
mn (t ) = ∑ p s ,n (t ) (1)
s =1

1 N
σ n (t ) =∑ ( p s,n (t ) − mn (t ))
N s =1
(2)

The selected features correspond to the times ti at which mn(ti) varies greatly for different
pocket depths while σn(ti) remains small. For a particular set of properties, ni, and their
corresponding times, ti, the feature vector for tooth site s would become
{
f s = p s ,ni (t i ) } (3)
In the end, 54 different features were selected from the DWFP properties from two
different mother wavelets.

Binary Classification

Once the features have been generated for the wavelet fingerprint properties, we
then apply various standard pattern classification techniques from the PRTools catalog of
MATLAB functions [20]. Many of these are Bayesian techniques, except in this case we
have set the prior probabilities as equal for all classes:
1. LDC: Linear discriminant classifier
2. QDC: Quadratic discriminant classifier
3. KLLDC: Linear classifier using Karhunen-Loéve expansion of covariance matrix
4. PCLDC: Linear classifier using principle component analysis
5. LOGLC: Linear logistic classifier
6. FISHERC: Linear minimum least square classifier
7. NMC: Nearest mean classifier
8. NMSC: Scaled NMC

1570
 9. POLYC: Untrained classification with additional polynomial features
10. SUBSC: Subspace classifier
11. KNNC: k-nearest-neighbor
The sampling method used for classification was a leave-one-out technique, in
which a single tooth site s was removed from the data set for testing, and the remainder
was used for training the classifier. Unfortunately, when the full classification matrix is
used with these classification maps, no less than 60% error is observed. In all of these
tests, the map tends to classify most of the waveforms in those pocket depths that have the
largest number of objects, namely, 2 and 3mm, because they are the most populated pocket
depths in the clinical data set. In order to counter this predisposition of the standard
classification schemes to classify all the objects into the highest volume classes, a binary
classification scheme was developed. The procedure is similar to the one-versus-one
technique of Support Vector Machines [21]. The basic idea is to classify the waveform
associated with any tooth site against only two possible classes at a time using standard
classifiers listed above. The training and test sets are divided from the data using the leave-
one-out technique. If the number of objects in each class differs, a random sample from the
larger class of equal size to the smaller class is chosen for training. This process is repeated
until at least 90% of the waveforms from the larger class size have been sampled. With
each repetition, the predicted labels are stored. The procedure is labeled a binary algorithm
because each classification is restricted to only two possible pocket depth values.
Dimensionality reduction is required to reduce the resulting array of labels into a single
value. In general, a majority voting method is used.

Classifier Combination

The binary classification algorithm as discussed above can be configured in many

different ways. For feature selection, we may choose to collapse the k index over the 30
different waveforms before or after feature selection, and the majority voting of the
dimensionality reduction can be performed in slightly different ways. All of these
methods were applied using eleven different maps listed above. The results of each
individual classifier was sorted by average accuracy within 1mm, and the highest percent
(10-95% accuracy) were combined using dimensionality reduction methods; the mean or
mode of the labels can be calculated (democratic voting), or the most probable label can be
calculated (majority voting). Combining classifiers can often reduce the error of pattern
classifiers but can never substitute for proper classification techniques applied when the
individual classifiers are formed [22].

RESULTS AND DISCUSSION

The binary classification algorithm was applied as described in the procedure

above. The primary method of measuring the success of each technique is finding the
percent of waveforms accurately described within 1mm per pocket depth and averaging
over the accuracy per pocket depth. If we instead tried to measure the total number of
waveforms accurately described within the manual probe's 1mm tolerance regardless of
pocket depth measurement, we tend to select for the classification techniques that
accurately describe only the most populated pocket depths. We performed these tests for
all 7 possible pocket depths from the manually measured data set, but we also restricted
the data sets to the 1-5mm pocket depths, since there are so few measurements in the 6-
7mm range in our patient population. We further restricted the possible pocket depths to
the 2-5mm set, since we felt the 1mm data sets might be poorly described by the ultrasono-

1571
TABLE 1. The accuracy (%) within the tolerance of 1mm for the most accurate binary classifiers and the
revised schemes that were manually selected for a larger spread of labels are shown in the table below.
Pocket Classifier Combination
Depths No Yes
Used [mm] Highest Accuracy Revised Highest Accuracy Revised
1-7 70.1 - 72 -
1-5 79.4 71.3 79.9 76.8
2-5 85.9 81 86.6 80.6

graphic probe because of overlapping reflections from inside the tip. We show here results
before and after classifier combination, and we show the highest average percent
accurately identified within 1mm as well as some revised selections that have a lower
accuracy but higher range of predicted labels. Table 1 gives the classification results in the
form of the average percent accurately identified within 1mm tolerance over the given
pocket depth range.
The results above show that the highest average percent accuracy for all the data
collected from 12 patients in the ODU clinical trials using a 5th generation prototype is at
best 86.6% within a tolerance of 1mm. Meanwhile, the best accuracy at that tolerance
using the 4th generation probe from previous methods without using pattern classification
was 60% for a single patient [14]. Note also that classifier combination does not
significantly increase accuracy.

CONCLUSION

We have described the development of the ultrasonographic periodontal probe and

the clinical tests on 12 patients. The resulting waveforms were analyzed with DWFP
processing and image recognition techniques applied to extract numerical measurements
from these wavelet fingerprints. These parameters were optimized and averaged to yield
training data sets for pattern recognition analysis, where testing data sets were configured
for a leave-one-out technique. The binary classification algorithm was described and
applied to these classification sets, and the labels from this technique were combined to
strengthen the results. Different sets of possible pocket depths can be used, since there
are small numbers of measurements in several of the pocket depths. The results can be
configured either to yield the highest average percent of waveforms correctly identified
within 1mm, or they can be configured to yield a larger spread in the type of labels with
approximately a 5% decrease in accuracy within 1mm. Overall, the results from the
classification scheme can identify ultrasonographic periodontal probe measurements
closely to the manual probe, so that 70.1-86.6% of the ultrasonographic measurements are
within 1mm of the manual measurement. These values are close to those in the literature
comparing other periodontal probes to the manual probe and may be due to the
imprecision and low reproducibility of the manual probe.
We conclude that the ultrasonographic periodontal probe is a viable technology and
that we can use sophisticated mathematical techniques to extract pocket depth information.
To yield better results, a larger training data set will be required.

ACKNOWLEDGEMENTS

The authors would like to acknowledge partial funding support from DentSply, Int'l
and from the Virginia Space Grant Consortium. We would also like to thank Gayle
McCombs of ODU for acquiring the clinical data and Jonathan Stevens for constructing
the investigational device.

1572
 REFERENCES

1. S. R. Ghorayeb, C. A. Bertoncini, and M. K. Hinders, IEEE Trans Ultrason Ferr Freq

Cont, 55, 1256–1266 (2008).
2. M. K. Hinders and J. A. Companion, “Ultrasonic periodontal probe,” in Review of
Progress in QNDE, 18b, edited by D. O. Thompson and D. E. Chimenti, AIP
Conference Proceedings, Plenum Press, NY (1998), pp. 1609–1615.
3. M. K. Hinders, A. Guan, and J. A. Companion, J Acoust Soc Am 104, p.1844 (1998).
4. J. A. Companion, “Differential measurement periodontal structures mapping system,”
US Patent No. 5,755,571 (1998).
5. M. K. Hinders, J. E. Lynch, and G. B. McCombs, “Clinical tests of an ultrasonic
periodontal probe,” in Review of Progress in QNDE, 21b, edited by D. O. Thompson
and D. E. Chimenti, American Institute of Physics, Melville, NY (2001), pp. 1880–
1890.
6. J. E. Lynch, “Ultrasonographic measurement of periodontal attachment levels,” Ph.D.
Thesis, College of William and Mary (2001).
7. J. E. Lynch and M. K. Hinders, Rev Sci Instrum 73, pp. 2686–2693 (2002).
8. J. Hou, “Ultrasonic signal detection and characterization using Dynamic Wavelet
Fingerprints,” Ph.D. Thesis, College of William and Mary, (2004).
9. F. A. Carranza and M. G. Newman, Clinical Periodontology, 8th edition, W B
Saunders Co, Philadelphia, PA (1996).
10. P. E. Peterson and H. Ogawa, J Periodontol 76, pp. 2187–2193 (2005).
11. N. P. Lang and E. F. Corbet, Int Dent J 45, pp. 5–15 (1995).
12. J. Hou and M. K. Hinders, Mate Eval 60, pp.1089–1093 (2002).
13. M. K. Hinders and J. R. Hou, “Ultrasonic periodontal probing based on the dynamic
wavelet fingerprint,” in Review of Progress in QNDE, 24b, edited by D. O. Thompson
and D. E. Chimenti, American Institute of Physics, Melville, NY (2004), pp. 1549-
1556.
14. J. Hou, S. T. Rose, and M. K. Hinders, Eurasip J on Appl Signal Processing 7, pp.
1137–1146 (2005).
15. K. Rudd, C. A. Bertoncini, and M. K. Hinders, Open Acoustics 2, pp.1–19 (2009).
16. J. E. Lynch, M. K. Hinders, and G. B. McCombs, Measurement 39, pp. 429–439
(2006).
17. M. K. Hinders and G. B. McCombs, Dim Dent Hygiene 4, pp. 16–18 (2006).
18. R. M. Haralick and L. G. Shapiro, Computer and Robot Vision, volume I, Addison-
Wesley, 1992, pp. 28-48.
19. B. K. P. Horn, Robot Vision, MIT Press, 1986.
20. R. P. W. Duin, “PRTools Version 4: A Matlab Toolbox for Pattern Recognition,” Delft
University of Technology (2007), URL http://www.prtools.org/files/PRTools4.1.pdf.
21. J. C. Platt, N. Cristianini, and J. Shawe-taylor, “Large margin dags for multiclass
classification,” in Advances in Neural Information Processing Systems, MIT Press,
2000, pp. 547–553.
22. L. I. Kuncheva, Combining Pattern Classifiers, Wiley, New York, 2004.

1573

View publication stats