2.1.

Plaque Collection Study

A plaque collection study was performed at the Technology Center Colgate-Palmolive Piscataway, New Jersey, in January 2016 to understand the autofluorescence characteristics of dental plaque. The study followed the instructions based on the Declaration of Helsinki (2008) and with the U.S. Code of Federal Regulations governing the protection of human subjects (21 CFR 50) and the institutional review boards (21 CFR 56). It was approved by Clinical Research Institutional Review Board, Concordia, New Jersey, under the registration number CRO-2015-1110-PLACOLL-JG.

Volunteers between 18 and 65 years old, having a minimum of 20 natural uncrowned teeth, and in good general health, were eligible to participate. They agreed to take part in this study by reading and signing an informed consent form and completed an oral hard and soft oral tissue exam, together with a general health questionnaire form.

Participants were excluded with advanced periodontal disease (purulent exudate, tooth mobility, and/or extensive loss of periodontal attachment or alveolar bone) and spontaneous bleeding of the gums. Immune compromised individuals (HIV, AIDS, immunosuppressive drug therapy), or subjects taking medications that may affect the oral flora (e.g., use of antibiotics 1 month prior to or during the study), were also excluded from the study. Further exclusion criteria were five or more carious lesions requiring immediate restorative treatment, tumor of the soft or hard tissues of the oral cavity, abnormal salivary function, premedication required for dental procedures, pregnant women, or women who were breastfeeding.

A total of 24 subjects were enrolled in the study. Participants abstained from oral hygiene at least for 24 h. During the experiment, dental plaque was collected from tooth surfaces by a certified hygienist and stored in a preweighted cylindrical test tube. Test tubes with plaque were then weighted again to determine the plaque weight. For analysis, 90% phosphate-buffered solution was added to each test tube to create a 10% plaque sample, and the tubes were vortexed for 1 min. The solution was then dispensed into a 96-well microplate and scanned by the Molecular Device SpectraMax M5 Spectrometer (Sunnyvale, California), first in excitation mode, and then in emission mode, to determine the optimal wavelength for generating a maximum fluorescence intensity response. For each plaque sample, the optimum excitation wavelength was determined from the wavelength giving the strongest emission intensity at 650 nm when scanning excitation spectrum from 300 to 600 nm. The optimum emission wavelength was defined as the strongest emission response from 500 to 800 nm generated using an excitation wavelength of 405 nm. The optimal excitation and emission wavelengths for each subject were recorded and used to calculate a mean and standard deviation across all subjects to inform the construction of an autofluorescence imaging system for plaque detection.

2.2.

Red Fluorescence Imaging System

Based on the excitation and emission experiment described above (results shown in Sec. 3.1), a fluorescence imaging equipment was developed at Colgate Palmolive Dental Health Unit, University of Manchester, for the acquisition of dental plaque images of human teeth. This consisted of two parts: a handheld imaging probe and an illumination controller.

The handheld probe allowed imaging of the full mouth. It was comprised of a lipstick color camera (IK-UM51H 1/3” Toshiba) fitted with a 17.5-mm focal length lens (58206 Edmund Optics), offering an effective field of view of $13.5\mathrm{mm}\times 16.5\mathrm{mm}$ able to cover the full range of a single incisor, canine, or molar. It is worth noting that the near-infrared filter, cutting around 650 nm, was removed from the lipstick camera to capture the red fluorescence signal in the spectrum beyond 650 nm.

The control box included a near-UV LED centered at 405 nm with power density of $37.1\mu \mathrm{W}/\mathrm{m}{\mathrm{m}}^2$ (M405L2–UV, Thorlabs Inc.) to provide the excitation light for red fluorescence imaging of dental plaque. During the clinical study, participants were provided with a pair of UV-protection glasses and dentist put on UV-protection gloves across the whole image acquisition process. The emission light was filtered by a 500 nm long-pass filter (62976, Edmund Optics), which was placed in front of the camera lens in the handheld probe.

A customized mirror tip, mounted on top of the handheld probe, allowed the full access into human mouth for imaging both anterior and posterior teeth. An air bump was connected to the imaging probe to clear the fogging on the mirror window caused by patient breathing during the measurement. In the clinical study, this mirror tip was replaced for every patient and every visit for infection control purposes.

Both the camera and illuminator were controlled by custom-written software. Camera settings, such as gain, shutter speed, image pixel resolution, etc., were preconfigured with a study design. This study design was locked unchangeable throughout the experiments. The brightness of the near-UV light was computer controlled by changing the output of a current controller that was connected to the LEDs. The light achieved ideal brightness when the average pixel intensity of a standard gray card reached a predefined value with a small tolerance of $\pm 0.05\text{pixel}$ intensity value. This made imaging illumination at different times comparable. During the image acquisition, the fluorescence images of human teeth were captured by moving and focusing the camera at baseline. For the follow-up visit, an edge detection overlay extracted from the baseline image was displayed on top of the live video. This overlay guided the dentists to reposition the imaging probe to capture the teeth areas as close as those from baseline, so that the plaque changes shown on the fluorescence images were comparable between visits. Figure 1 shows the red fluorescence imaging system stated above.

Fig. 1

Red fluorescence imaging system. (a) The whole imaging system, including imaging probe, control box, and custom software. (b) Imaging probe with excitation light turned on. Blue light reflected by the disposable intraoral tip. (c) Imaging probe applied during clinical study.

2.3.

Plaque Detection Using Red Fluorescence Imaging

2.4.

Automatic Plaque Detection and Classification on Fluorescence Images

2.4.1.

Automatic plaque detection

Associating autofluorescence characteristics with the spectral responses of conventional RGB cameras, plaque areas present apparent higher intensity in the red channel while the healthy teeth fluoresce green due to the intrinsic property of hard dental tissues. Dental plaque distribution ${\mathrm{D}}_{\mathrm{p}}$ therefore can be revealed by the quotient between two spectral bands, as defined in Eq. (2). According to the Lambertian law,¹³ this division cancels out the shading effects due to tooth curvature:

Eq. (2)

$${\mathrm{D}}_{\mathrm{p}}=\frac{{\mathrm{I}}_{\text{Red}}}{{\mathrm{I}}_{\text{Green}}},$$

where ${\mathrm{I}}_{\text{Red}}$ and ${\mathrm{I}}_{\text{Green}}$ stand for the red and green channels of the fluorescence image, respectively.

A binary mask ${\mathrm{M}}_{\mathrm{p}}$ can be identified by setting threshold in the plaque feature space ${\mathrm{D}}_{\mathrm{p}}$. Pixels having intensity above the threshold were classified as plaque and labeled 1, and the rest of the pixels were considered as healthy dental tissue marked 0. In this study, the plaque feature was regulated to an interval of 0 to 5. The threshold was set to 1.0 as a constant throughout the work, which means 20% of the intensity range in the feature space ${\mathrm{D}}_{\mathrm{p}}$.

For a selected tooth outlined by the dentist, the average ${\mathrm{D}}_{\mathrm{p}}$ value of nonplaque pixels was considered as the intensity level of sound adjunct enamel. Therefore, a plaque intensity measure $\mathrm{\Delta }P$ can be calculated for each plaque pixel within the chosen tooth:

Eq. (3)

$$\mathrm{\Delta }P(x)={\mathrm{D}}_{\mathrm{p},{\mathrm{M}}_{\mathrm{p}}==1}(x)-\overline{{\mathrm{D}}_{\mathrm{p},{\mathrm{M}}_{\mathrm{p}}==0}},$$

where $x$ represents a plaque pixel, and $\overline{{\mathrm{D}}_{\mathrm{p},{\mathrm{M}}_{\mathrm{p}}==0}}$ is the average intensity of sound adjunct enamel in the plaque feature space.

A three-dimensional feature vector, including the average plaque intensity $\mathrm{\Delta }{P}_{\mathrm{ave}}$, the area percentage of plaque pixels within the whole tooth $\delta$, and their dot product ${\mathrm{\Delta }P}_{\mathrm{ave}}*\delta$, was calculated. These features were then input into an ensemble classifier as the diagnostic descriptors to estimate the plaque level on each tooth.

Figure 2 shows the red fluorescence images scored 0 to 3 by the certified dentist at tooth level with the detected plaque areas overlaid.

Fig. 2

Red fluorescence images scored 0 to 3 by the examiner at tooth level with detected plaque areas displayed. (a)–(d) Red fluorescence images scored 0 to 3 (from top to bottom) by a certified dentist. (e)–(h) Dental plaque distribution ${\mathrm{D}}_{\mathrm{p}}$ of the selected tooth, where green/yellow color refers to small ${\mathrm{D}}_{\mathrm{p}}$ value and orange/red colour stands for large one. (i)–(l) Detected plaque areas (blue color) within the selected tooth overlaid on top of fluorescence images.

2.5.

Experiments and Statistic Analysis

In this pilot study, a number of 25 subjects were screened, of which 20 met the inclusion criteria and 15 were enrolled after informed consent. One participant dropped out at 24 h visit during the study. Hence, this paper reported the results based on the 14 participants who followed the entire protocol.

Excluding the scores of “cannot access,” there were 756 clinical disclosing plaque scores and 724 dentist RFT-QH scores included in the experiment. A threefold cross-validation¹⁸ was combined with the ensemble classifier as the training and testing strategy to derive the computer red fluorescence plaque scores at tooth level.

The performance of plaque classification was evaluated using confusion matrix. Sensitivity (true positive number/number of plaque cases), specificity (true negative number/number of healthy cases), and the total agreement ($\text{true}\text{positive number}+\text{true negative number}/\text{total number}$)¹⁹ were also recorded. A linear weighted kappa²⁰ was performed to measure the level of agreement between dentist diagnosis and automatic classification. Differences of computer features (plaque intensity, plaque areas, and their product) between plaque levels (RFT-QH) were calculated by $t$-test, and those between different visits were analyzed by paired $t$-test.

Spearman rank correlation²¹ was used to determine the association between disclosing plaque indices with red fluorescence RFT-QH scores and computer features. Differences of plaque scores between visits were calculated by Wilcoxon signed-rank test.

Based on the results derived from this pilot study, a post-hoc power analysis²² was performed to outline the sample size for future clinical trials. Power analysis was conducted using the distributions of computer plaque scores for differentiating plaque level at baseline from that after 48 h plaque accumulation.

The automatic plaque detection and classification algorithms were implemented using C++ under Visual Studio 2010 (Microsoft, United States), running on a platform of Intel i7-4770 at 3.40 GHz CPU and 8GB DDR3-1600 RAM. Statistical analysis was performed using SPSS (version 23, IBM Inc., United States), and post-hoc power analysis was conducted using R.

3.1.

Excitation and Emission Spectrum of Dental Plaque

From Fig. 3, it is noted that there was little variation of the excitation and emission spectrum of dental plaque in peak wavelength among subjects. The average excitation peak of 24 subjects was $403.79\pm 0.59\mathrm{nm}$ and the average emission peak was $639.75\pm 1.26\mathrm{nm}$. This was consistent to the previous findings, which claimed that porphyrins have the excitation band around 400 nm and the emission wavelength after 600 nm.^7,8

Fig. 3

Excitation and emission spectrum of dental plaque from 24 subjects measured by spectraMax M5 spectrometer. Colors represent fluorescence characteristics of plaque from different subjects. (a) Excitation spectrum from 350 to 600 nm in excitation mode. (b) Emission spectrum from 500 to 750 nm in emission mode generated using an excitation wavelength of 405 nm.

3.2.

Automatic Classification Accuracy of Red Fluorescence Plaque

Table 2 shows the automatic classification results of red fluorescence plaque derived from the proposed methods taking the dentist RFT-QH assessment as the reference standard. The total classification accuracy was 62.7%, where the true positive rates of plaque level 0 to plaque level 3 are 67%, 50%, 68.9%, and 79.1%, respectively. A moderate agreement showed between dentist RFT-QH scores and computer RFT-QH scores with a linear weighted kappa of 0.594.

Table 2

Automatic classification results (number/percentage) of red fluorescence plaque at tooth level using RUSBoost classifier with threefold cross-validation. Values in the left diagonal cells represent the true-positive number and rate in each class. Values in other cells show the number and rate of misclassified cases of different plaque levels.

Figure 4 demonstrates the distributions of computer features at different plaque levels (RFT-QH) for both dentist assessment and computer classification. Plaque intensity ($\mathrm{\Delta }{\mathrm{P}}_{\mathrm{ave}}$) increased and the plaque areas ($\delta$) enlarged with the increment of RFT-QH scores. Taking the significant level at $p<0.05$, all the computer features resulted in significant differences between any two plaque levels within the RFT-QH scale.

Fig. 4

Boxplot describing the distributions of computer features at different red fluorescence plaque levels. Red line in each box represents the median value of the feature in each class. Box boundaries in blue lines show the 25% quartile (lower) and 75% quartile (higher) of the data. Boundaries in black lines describe the extreme values in the tails of the feature distribution to rule out the outliers. We set whisker parameter to 1 in this case. Red cross beyond black lines is the outlier. (a)–(c) Distribution of computer features based on dentist plaque assessment. (d)–(f) Distribution of computer features based on computer plaque classification. (a) and (d) Average plaque intensity $\mathrm{\Delta }{P}_{\mathrm{ave}}$. (b) and (e) Plaque areas $\delta$ (%). (c) and (f) Inner product of plaque intensity and plaque areas $\mathrm{\Delta }{P}_{\mathrm{ave}}*\delta$.

3.3.

Correlation Between Clinical Disclosing Plaque and Red Fluorescence Plaque

From Table 3, a moderate to strong correlation was identified between clinical disclosing plaque and red fluorescence plaque at subject level. Tooth level association was overall lower presenting weak to moderate correlations. Plaque areas ($\delta$) and inner product ($\mathrm{\Delta }{P}_{\mathrm{ave}}*\delta$) correlated well with dentist RFT-QH scores at tooth and subject levels throughout the experiment. Their correlations with clinical assessment were moderate to strong at subject level and weak to moderate at tooth level (Table 4). This observation was consistent to the results between clinical plaque scores and RFT-QH scores based on the red fluorescence images. The average plaque intensity ($\mathrm{\Delta }{P}_{\mathrm{ave}}$) was less relevant with human visual assessment giving moderate correlations at most.

Table 3

Spearman rank correlation between clinical disclosing plaque scores and red fluorescence RFT-QH index at tooth and subject levels. Correlations without superscript “*” are significant at 0.01 level (two-tailed).

Table 4

Spearman rank correlation between computer features and human visual assessment at tooth and subject levels. Correlations without superscript “*” are significant at 0.01 level (two-tailed).

Considering red fluorescence score $\mathrm{RFT}-\mathrm{QH}>=1$ as plaque affected, Fig. 5 demonstrates the agreement between clinical disclosing plaque indices and red fluorescence plaque scores, taking the disclosing parameters as assessment references. It is noted that, with the increase of threshold in disclosing T-QH, sensitivity of plaque detection by the red fluorescence imaging increased while the corresponding specificity decreased. The agreement between dentist disclosing scores and computer RFT-QH scores presented comparable results with that between disclosing scores and dentist RFT-QH assessment. The best agreement of plaque detection was observed, when disclosing plaque threshold at level 2, for both dentist RFT-QH evaluation (sensitivity 71.1%, specificity 67.7%, agreement 70.2%) and computer RFT-QH classification (sensitivity 68.4%, specificity 62.9%, agreement 67.1%).

Fig. 5

Agreement between clinical disclosing plaque scores and red fluorescence plaque scores (RFT-QH). (a) Disclosing plaque scores versus dentist RFT-QH scores. (b) Disclosing plaque scores versus computer RFT-QH scores.

3.4.

Plaque Progression Between Visits

Figure 6 shows tooth level plaque progressions described by the dentist diagnosis and computer analysis during the clinical study. Gradual increase was found in human visual assessment of dental plaque. Similarly, the average values of computer features increased with the visit number. Taking significant level at $p<0.05$, statistically significant differences were detected between baseline, 24 h, and 48 h visits in the clinical disclosing T-QH scores, dentist RFT-QH scores, and computer RFT-QH scores at tooth and subject levels. For computer features, plaque areas ($\delta$) showed significant differences between any consecutive visits. But there was no significant difference in plaque intensity ($\mathrm{\Delta }{P}_{\mathrm{ave}}$) between baseline and 24 h ($p=0.076$) as well as 24 h and 48 h ($p=0.07$) at tooth level and between 24 h and 48 h ($p=0.15$) at subject level.

Fig. 6

Tooth level plaque progressions described by dentist plaque scores, computer plaque scores, and computer features during the clinical study. (a) Case number of disclosing T-QH scores at different levels across the visits. (b) Case number of dentist RFT-QH scores at different levels across the visits. (c) Case number of computer RFT-QH scores at different levels across the visits. (d) Average values of computer red fluorescence features at different visits.

Because pooling cases (excess fluorescence dye not completely dried) are likely to be scored $\mathrm{T}-\mathrm{QH}=1$ but less clinically relevant, this study took the disclosing score $\mathrm{T}-\mathrm{QH}\ge 2$ and red fluorescence index $\mathrm{RFT}-\mathrm{QH}\ge 1$ as plaque affected. Figure 7 shows plaque progressions in terms of affected tooth number per subject by human visual evaluation and computer classification. It is noted that the affected tooth number per subject increased with time for both disclosing scores and red fluorescence scores. After 48 h with no oral hygiene, all the subjects had six teeth affected with plaque, which was detectable by disclosing solutions in clinical assessment. Red fluorescence plaque was not identified on all the tooth surfaces. But for subjects at 48 h appointment, there were 10 out of 14 participants having at least five teeth affected with red fluorescence plaque by dentist diagnosis. Computer analysis showed lower plaque level, giving eight subjects in this case.

Fig. 7

Plaque progressions in terms of affected tooth number per subject evaluated by (a) clinical disclosing plaque, (b) dentist red fluorescence plaque, and (c) computer red fluorescence plaque.

3.5.

Post-hoc Power Analysis

A post-hoc power analysis was performed to outline the sample size in the future clinical studies for differentiating plaque level at baseline from that after 48 h plaque accumulation, based on the distributions of computer plaque scores derived from this pilot study. Taking the significant level $p<0.05$ (Wilcoxon signed-rank test, two tails) and statistical power $>0.9$, the sample size is minimum 72 at tooth level with effect size of 0.42 and minimum 42 at subject level with effect size of 0.56.