Pattern Classification Techniques for the Ultrasonographic Periodontal Probe Crystal Bertoncini Nondestructive Evaluation Laboratory The College of William and Mary Introduction In the clinical practice of dentistry, radiography is used to detect subsurface structural defects such as cavities and cracks. However, radiation is harmful to the patient, and has been shown to actually lead to cavities by the demineralization of teeth. Radiography can also only detect defects parallel to the projection path, and it is useless for detecting conditions such as gum disease because soft tissues are transparent to x-rays. Ultrasonographic Nondestructive Evaluation, however, is safe to use as often as indicated, and computer interpretation software makes diagnosis automatic. The structure of soft tissues can be analyzed effectively with ultrasound, and even symptoms such as inflammation can be registered with this technology [1]. One application of ultrasound (US) in dentistry, a periodontal probe, is already being developed as a spin-off of NASA technology [2]. Periodontal disease is caused bacterial infections in plaque, and the advanced stages can Figure 1: Manual vs US Probing: The figure compares a) manual periodontal probing, in which a thin metal probe is inserted in between the tooth and gums, versus b) the ultrasonographic periodontal probe, in which ultrasound energy propagates through water noninvasively. cause tooth loss when the periodontal ligament, which usually holds the tooth in place, erodes. The usual method of detection is with a thin metal probe lined with gradations marking depth in millimeters (Fig. 1a). The dental hygienist inserts the probe into the area between the tooth and gum to measure the depth to the periodontal ligament. This method is only accurate to +/- 1 mm 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. The ultrasonic periodontal probe currently under development uses high-frequency ultrasound to find the depth of the periodontal ligament non-invasively. An ultrasonic transducer projects high frequency (10-15 MHz)
ultrasonic energy in between the tooth and the gum and detects echoes of the returning wave (Fig 1b). Since water is used as a coupling agent, the depth of measurement can be calculated by multiplying the time delay of the echoes by the speed of sound in water (1500 m/s). Figure 2: Experimental Apparatus: All devices necessary to conduct experiments on the ultrasonic periodontal probe are shown, including a) laptop computer, b) interface device, c) pulser-receiver, d) ultrasonic probe, e) manual probe, and f) foot pedal to control water flow. Clinical Trials and Results Clinical trials were performed between April and May of 2007 with the assistance of Gayle McCombs at ODU s Dental Research Center using a 5 th generation prototype ultrasonogrpahic periodontal probe on 12 patients (Fig 2) 1. The dental hygenists operated the probe while the computer interface was monitored to assure that the data was of high quality. For the next year and a half, artificial intelligence (AI) algorithms were developed to automatically interpret the ultrasonic echoes in real time. The signal-analysis technique called Dynamic Wavelet Fingerprinting (DWFP) was used to develop a pattern classifier that determines the pocket depth [3]-[4]. This technique applies a wavelet transform on the original signal, which results in loop features that resemble fingerprints (Fig. 3), but classifying those features required an innovative classification technique. The Pattern Classifier for the Ultrasonographic Periodontal Probe To create the pattern classifier, the waveforms from the clinical trials were fingerprinted using DWFP, and properties of those fingerprints were measured using image recognition techniques. These properties included fitting an ellipse matching the second moments of the fingerprint and capturing elements such as eccentricity, inclination, and height. In all, 22 different properties were recorded from each of two different mother wavelets. Features were selected from this large group to use in classification by selecting values of the fingerprint properties at a time in which the average of the fingerprint property varied between the different 1 The author acknowledges funding provided by DentSply, International.
possible pocket depths. However, standard pattern recognition techniques resulted in labeling all of the waveforms in the post populated pocket depth range, 2-3 mm, since that yielded the most correct label assignments. Binary Classification Algorithm In order to develop a classifier that would even consider the larger pocket depth values, a binary classification algorithm was developed. The basic idea is to apply basic pattern classifier maps to classify the waveform associated with any tooth site against only two possible classes at a time. 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 the classification. 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. To use most of the data for training, the algorithm proceeds by the leave-oneout technique, so that one waveform is extracted from the data set while the rest is used to train the classifier. The single remaining waveform is used to test the results. However, this results in a large-dimensional label matrix. To reduce the dimensions, several techniques are applied, including finding the mode of the predicted labels and finding a weighted probability. Lastly, to maximize the accuracy of the classifier, many different classifier maps were used and the results combined. After the results were gathered, the classification schemes with the best accuracy were not necessarily optimal. Instead, revised classification schemes with a larger spread of predicted labels but lower accuracy were selected. Figure 3: Dynamic Wavelet Fingerprinting Technique: Shown is a) the original waveform, b) the filtered waveform windowed in on an area of interest, and c) the filtered signal after a wavelet transform has been applied. The primary method of measuring the success of each technique is finding the percent of waveforms accurately described within 1 mm per pocket depth and averaging over the accuracy per pocket depth. The best results are summarized in Table 1. Classification schemes were configured for all 7 possible pocket depths from the manually measured data set, but the data sets
were also restricted to the 1-5 mm pocket depths, since there are so few measurements in the 6-7 mm range in our patient population. The possible pocket depths were further restricted to the 2-5 mm set, since the 1 mm data sets might be poorly described by the US probe because reflections from inside the tip seem to cover that range. The columns indicate whether or not several classifiers were combined and whether or not a revised classification scheme was selected to increase the spread of labels. Table 1: Summary table of accuracy within 1 mm (%). The table below shows accuracy (%) within the tolerance of 1 mm for the most accurate and revised classification schemes, which involve more widely-spread labels than the original. Current and Future Research Pocket Depths Classifier Combination Used (mm): No Revised Yes Revised 1-7 70.1-72.0-1-5 79.4 71.3 79.9 76.8 2-5 85.9 81 86.6 80.6 To truly test the robustness of the binary classification algorithm for the ultrasonographic periodontal probe, much more data is required. A proposal is currently submitted to acquire 500 more patients to test and possibly redesign the classifier. In the intervening time, it may be more interesting to test the binary classification algorithm on other cases, including developing an ultrasonic probe to detect cavities in teeth, finding flaws in pipes [5], and predicting roof falls in mines [6]. The dental industry is currently interested in developing technology that can detect cavities or other flaws that occur in between teeth, which x-rays are unsuited to identify. However, ultrasonic guided waves, which are surface waves that can travel around curves, may be able to detect these hard-to-find cavities. Preliminary results using an ultrasonic thickness gauge show promise and a thousand clean, extracted teeth will be sent to our laboratory for testing. The reflections from the cavities are small and difficult to differentiate without the pattern classification techniques described above. Another application of the pattern classification techniques developed for the periodontal probe is detecting flaws in pipes. Ultrasonic guided waves are uniquely suited to detect flaws around bends in pipes, but diffraction and interference increase the difficulty of resolving these flaws. Pattern classification techniques may be useful to indicate the presence of flaws in these complex reflections.
Lastly, roof fractures in mines can occur with very little warning. Previous publications [6] attempted to predict roof falls in mines by recording echoes from 15 different ultrasonic transducers placed in the mine. Preliminary work indicates that precursor activity may be measurable in many of the cases of roof falls. The pattern classification techniques developed here may be able to improve the accuracy of predicting roof falls from these precursor events. Space Applications All of these devices use analysis techniques that are advantageous to the space industry. However, a more direct application of these devices to the space industry are the possible dental technologies themselves. In long-term flight situations that span months or years, a physician is not always available, and it is extremely unlikely that a dentist will be available should dental problems occur. The diagnostic techniques described above can be used to identify caries, surface fissures, tooth decay, and periodontal disease in other words, they span the range of almost any dental problem that an astronaut can experience both pre-flight and during long-term flight situations. References [1] SR Ghorayeb, CA Bertoncini, and MK Hinders (2008) Dental Ultrasonography, IEEE Trans Ultrason Ferr Freq Cont 55(6): 1256-1266. [2] Differential measurement periodontal structures mapping system, United States Patent 5755571, Companion, John A., The United States of America as represented by the Administrator NASA] [3] JD Hou and MK Hinders (2002) Dynamic wavelet fingerprint identification of ultrasound signals, Mate Eval, 60(9):1089-1093. [4] JD Hou, ST Rose, and MK Hinders (2005) Ultrasonic periodontal probe based on the dynamic wavelet fingerprinting, EURASIP J App Signal Processing, 7:1137-1146. [5] KE Rudd, KR Leonard, JP Binham, MK Hinders (2007) Simulation of guided waves in complex piping geometries using the elastodynamic finite integration technique, J. Acoust. Soc. Am, 121:1449-1458. [6] AT Iannacchione, T Batchler, TE Marshall (2004) Mapping hazards with microseismic technology to anticipate roof falls a case study, Proc. 23 rd Int l Conf. on Ground Ctrl in Mining, Moarganton, WV, 327-333.