1.

Introduction

Dental caries is a dynamic disease characterized by tooth demineralization leading to an increase in the porosity of the enamel surface—commonly known as white spots due to the white appearance created by the increased refractive index gradient.¹ Leaving these lesions untreated can potentially lead to cavities reaching the dentin and pulp and eventually causing tooth loss. Occlusal and approximal surfaces are among the most susceptible sites of demineralization due the acid attack from bacterial by-products in the biofilm.² The use of preventive agents to inhibit, or reverse the demineralization process is predicated on the detection of lesions at an early stage.^{3, 4, 5, 6} However, detecting early lesions is a difficult diagnostic task and quantitative techniques are still required particularly to monitor the caries lesion progression or arrest and therefore to assist in the decision making.

Noninvasive detection of white spots relies on the estimation of surface porosity or mineral loss. Although radiographic methods are suitable for approximal surface lesion detection, they offer a reduced utility for screening early caries in occlusal surfaces due to their lack of sensitivity at very early stages of the disease.^{7, 8, 9} Electrical caries monitoring has demonstrated good performance for single-point measurements;^{10, 11} however, imaging methods have the advantage of being more illustrative and allow relative measurements between different sites in the same tooth. Optical methods have been exploited due to their nondestructive nature and ability to detect early demineralization.^{12, 13}

New potential methods include polarized Raman spectroscopy,¹⁴ multiphoton imaging, infrared tomography, infrared fluorescence, optical coherence tomography (OCT), and terahertz imaging.¹⁵ Current imaging methods are based on the observation of the changes of the light transport within the tooth, namely absorption, scattering and/or fluorescence of light. Porous media scatters light more than uniform media and stain tends to absorb the light. Fiber-optic transillumination (FOTI) is a qualitative method used to highlight the lesions within teeth by observing the patterns formed when white light, pumped from one side of the tooth, is scattered away and/or absorbed by the lesion.¹⁶ DiFOTI^™ is a digital technique that is based on the transillumination of light through teeth for the assessment of dental caries.¹⁷ This technique is however difficult to quantify due to an uneven light distribution inside the tooth. This drawback is partially overcome with quantitative light fluorescence (QLF), an imaging method that relies on the natural fluorescence of teeth.^{18, 19} This fluorescence acts as an internal source of light that will try to escape through the surface of the tooth. With appropriate filters, one can observe the fluorescent light and can quantify the loss of mineral by visualizing the dark patches produced by scattering and/or absorption of light. QLF is a promising technique; however it has a downfall when trying to discriminate between white spots and stain because both produce the similar effect (extinction of light shadows). Stain is commonly observed in the occlusal sites of the teeth, and this obscures the true detection of caries. Stain therefore is one of the most crucial confounding factors in the detection of early caries lesions.

It has been shown however that working in the near-infrared (NIR) region of the spectrum can overcome the difficulties found with the imaging methods because scattering in enamel is reduced and absorption by stain is low.^{20, 21, 22, 23, 24} In fact, it has been demonstrated that the scattering by enamel tissues reduces in the form of $1∕{\lambda }^3$ as reported for laser wavelengths, $\lambda$ , of 512, 632, and $1053\mathrm{nm}$ .²⁰ In addition, a higher transparency for $1310\mathrm{nm}$ than for $1550\mathrm{nm}$ in sound enamel has been found, suggesting that water in the enamel attenuates the light at higher wavelengths.²¹ This has been used in optical coherence tomography for structural imaging of teeth and in tooth transillumination for interproximal and occlusal caries screening.²⁵ It has been demonstrated that stimulated lesions in tooth samples up to $6.75\mathrm{mm}$ in thickness can be resolved with a contrast ratio of $>0.35$ between sound and demineralized enamel by using $1310\mathrm{nm}$ .²² Although NIR imaging has shown a potential for delimiting carious regions successfully, not much work has been done to quantify the degree of lesions in naturally decayed teeth.

Our aim is to show for the first time that spectral imaging in the NIR region could be used to map and quantify the lesions from occlusal tooth surfaces by using the spectral signatures of water absorption and the effect of porosity in the scattering of light.

2.

Materials and Methods

2.2.

Instrumentation

An NIR hyperspectral camera system (Spectral Camera SWIR, ImSpector, Finland) was used to capture the spectral reflectance from the samples. The instrument images a line of vision at a time and diffracts the light onto a two-dimensional mercury cadmium tellaride (MCT) sensor array using a diffraction grating. A complete stack of spectral images (spectral data cube) is obtained by translating the sample at constant speed and line-imaging synchronously. The spectral analysis range was from $1000\text{to}2500\mathrm{nm}$ with a spectral resolution of $10\mathrm{nm}$ . Figure 1 shows an image of the system setup. Note that a two-side illumination using halogen lamps was configured to reduce the shadowing effects.

Fig. 1

NIR hyperspectral imaging system. For our investigations, the targets were teeth having different degrees of lesion.

Calibrated spectral reflectance data were obtained using the system’s associated software (ENVI 4.3 + in-build ImSpector add-ons). The calculation required to input a “raw” spectral data cube file, a dark current measurement associated with the data acquired (for noise subtraction), the reflectance spectra of a reference object, and its associated dark current measurement. The following calculation was performed for each pixel:

Eq. 1

$R\left(\lambda \right)=\frac{R_{\mathrm{r}}\left(\lambda \right)-D_{\mathrm{r}}}{W_{\mathrm{ref}}\left(\lambda \right)-D_{\mathrm{w}}}{W}_{\mathrm{spec}}\left(\lambda \right).$

Here $R\left(\lambda \right)$ , $R_{\mathrm{r}}\left(\lambda \right)$ , and $W_{\mathrm{ref}}\left(\lambda \right)$ are the calibrated, raw and white reference measurements at wavelength $\lambda$ , respectively. $W_{\mathrm{spec}}\left(\lambda \right)$ corresponds to the white reference reflectance at wavelength $\lambda$ as specified by the manufacturer. In addition, $D_{\mathrm{r}}$ and $D_{\mathrm{w}}$ are the dark current noise measurements obtained for the raw and white reference reflectance data.

2.4.

Image Analysis

As an example, a selection of the spectral images obtained from an occlusal tooth surface is shown in Fig. 2 at the wavelengths indicated. It is possible to see that the reflectance intensity drops at higher wavelengths. A typical reflectance spectrum for sound, white-spot (enamel lesion), and a dentin lesion areas is shown in Fig. 3 . Well-defined signatures can be observed among the three cases. For instance, the reflectance above $1400\mathrm{nm}$ is clearly higher for the decayed areas and, in particular, at $1610\mathrm{nm}$ . However, the absorption dips are further pronounced for the case of dentin lesion and especially at $1440\mathrm{nm}$ .

Fig. 2

Selection of spectral images. On the top left, a color picture of the tooth is shown. The remaining reflectance images are obtained with the NIR hyperspectral camera at the wavelengths indicated. (Color online only.)

Fig. 3

Example of NIR spectral reflectance curves from an occlusal tooth surface. The example shows the reflectance for a sound (black ◆), an enamel lesion (blue ∗), and a dentin lesion (red ○) region. The inset picture shows the location of the points selected within the tooth for this example. Vertical dotted lines indicate the wavelengths chosen for the example in Fig. 2. (Color online only.)

The spectral intensity at each pixel was normalized to the reflectance obtained at $1090\mathrm{nm}$ because it is the wavelength with the highest reflectance (least extinction). The degree of enamel, $S_{\mathrm{e}}$ , and dentin, $S_{\mathrm{d}}$ , lesions were therefore calculated as follows:

Eq. 2

$S_{\mathrm{e}}=\frac{R\left(1610\mathrm{nm}\right)}{R\left(1090\mathrm{nm}\right)},$

Eq. 3

$S_{\mathrm{d}}=\frac{R\left(1610\mathrm{nm}\right)-R\left(1440\mathrm{nm}\right)}{R\left(1090\mathrm{nm}\right)}.$

A caries score, $S_{\text{caries}}$ , was calculated as a combination of $S_{\mathrm{e}}$ and $S_{\mathrm{d}}$ and was designed to account for the deepest lesion observed by the NIR spectrum as follows:

Eq. 4

$S_{\text{caries}}=p\left(\frac{S_{\mathrm{e}}-K_{\mathrm{e}}}{N_{\mathrm{e}}}\right)+(1-p)(1+\frac{S_{\mathrm{d}}-K_{\mathrm{d}}}{N_{\mathrm{d}}}).$

Here, $K_x$ and $N_x$ are the enamel $(x:e)$ and dentin $(x:d)$ score calibration offset and normalization factor, respectively. In addition, $p=1$ if $[1+(S_{\mathrm{d}}-K_{\mathrm{d}})∕N_{\mathrm{d}}]<S_{\mathrm{dth}}$ , where $S_{\mathrm{dth}}$ represents the dentin lesion threshold, otherwise $p=0$ . Note that $S_{\text{caries}}$ depends on either but not both scores simultaneously. $S_{\mathrm{e}}$ is linked to the enamel scattering through the spectral signature sitting around $1610\mathrm{nm}$ , whereas $S_{\mathrm{d}}$ is linked to the water absorption normally in carious dentin through spectral signature sitting around $1440\mathrm{nm}$ ; the strongest effect is the one accounted for by $S_{\text{caries}}$ .

Outliers and noise introduced by specular reflections were removed by limiting the values of the numerators in Eq. 4 to the range $0<(S_{\mathrm{e}}-K_{\mathrm{e}})<M_{\mathrm{e}}$ and $0<(S_{\mathrm{d}}-K_{\mathrm{d}})<M_{\mathrm{d}}$ ; here $M_{\mathrm{e}}$ and $M_{\mathrm{d}}$ denote the upper limits. Values outside these limits were set to zero.

To fix the range of $S_{\text{caries}}$ from 0 to 2, the normalization factors in Eq. 4 were expressed as $N_{\mathrm{e}}=M_{\mathrm{e}}∕S_{\mathrm{dth}}$ and $N_{\mathrm{d}}=M_{\mathrm{d}}$ . Therefore, for values in the range of $S_{\mathrm{eth}}<S_{\text{caries}}⩽S_{\mathrm{dth}}$ , where $S_{\mathrm{eth}}$ represents the enamel lesion threshold, the score indicates an enamel lesion and for $S_{\text{caries}}>S_{\mathrm{dth}}$ , the score indicates a dentin lesion. Note that sound areas have a value of $S_{\text{caries}}⩽S_{\mathrm{eth}}$ .

The calibration values for the equations above were obtained empirically. For our case, results were best described with $S_{\mathrm{dth}}=0.05$ , $S_{\mathrm{dth}}=1.1$ , $K_{\mathrm{e}}=0.14$ , $K_{\mathrm{d}}=0.05$ , $M_{\mathrm{e}}=0.35$ , and $M_{\mathrm{e}}=0.15$ .

3.

Results

Twelve teeth with different lesion degrees were imaged and processed following the method described in Section 2.5. As an example, false-color-coded $S_{\text{caries}}$ maps for four teeth are shown in Fig. 4 . The associated color pictures and the selected histological sections chosen are also shown.

Fig. 4

Examples of teeth with different carious lesions. (a) Images show the color picture for each tooth, (b) Images show the corresponding histological section indicated by the red line in the color picture with the red arrow indicating the point of view of the section. (c) Images show the false-color NIR caries maps, spatially scored using $S_{\text{caries}}$ , for each tooth. (Color online only.)

Note that the extension of the lesion is not evident from the color pictures (which rely on visible wavelengths). It is possible to see however that the caries score maps clearly depict the lesion spatial distribution. This is possible due to the longer light penetration of NIR wavelengths through enamel. In addition, the depth of the lesions can be confirmed from the histological sections. The red arrows pointing toward the color picture indicate the point of view of the histological section, which is indicated by the red line.

For the examples presented in Fig. 4, lesion in tooth A rest within the enamel, having $S_{\text{caries}}<1.1$ , whereas the lesions in teeth B, C, and D reach the dentin, having $S_{\text{caries}}>1.1$ . This observations can be confirmed with the histological sections shown and scored as EDJ for tooth A and D2 for teeth B, C, and D. The black points highlighted in each image correspond to the maximum scores found within the region corresponding to the histological section and the average value is indicated by the arrow.

Hidden lesions as the one in tooth D are of particular interest since visual inspection does not reveal them easily and they could be left untreated increasing the risk of tooth loss. It is clear from the histology section that the damage has reached the dentin for this case and this is adequately indicated by values of $S_{\text{caries}}>1.1$ around the decayed region.

Detection of caries in pits and fissures is of great interest among the dental practitioners because demineralization commonly starts in these sites and the detection of them is not always obvious, especially in the presence of stain. The decay in the fissure pattern of tooth A can be easily discriminated in the $S_{\text{caries}}$ map; the demineralized part of the U-shaped fissure of this tooth is well described and confirmed with the histological section. Enamel demineralization in the fissure pattern of the remaining teeth shown in Fig. 4 are also depicted by their corresponding $S_{\text{caries}}$ map, but the deepest tooth lesion is the one considered for our statistical analysis below.

Areas in which there has been complete loss of tooth tissue creating a cavity (i.e. a hole), as the ones appearing in the right side of teeth B and C, yield low values for $S_{\text{caries}}$ . We believe this is caused by insufficient mineral and water to create a significant change in the spectrum. More investigations are needed to see the effects of exposed dentin when scoring the tooth with this method.

The NIR images are affected by specular reflection and particularly around the edges and the crests of the teeth, where reflections are expected to be strong. These reflections act as a noise factor, especially for the calculation of $S_{\mathrm{e}}$ in Eq. 2. Note that the stain and pigmentation in all teeth did not show a significant interference with the measurements.

It is important to emphasize that a direct relationship between histological depth and the magnitude of $S_{\text{caries}}$ has yet to be determined; however, it can be seen that the NIR algorithm discriminates well between sound, enamel, and dentin lesions.

A histogram showing the distribution of samples according to the different histological scores is shown in Fig. 5 . Note that three selected samples had a representative histological sound slice, four had a slice with an enamel lesion, and five had a slice with a dentin lesion.

Fig. 5

Sample distribution according to their histology score.

Figure 6 shows a correlation graph between the two scores and a Spearman’s correlation 0.89 significant at a level $p<0.01$ was found. In addition, a sensitivity of $>99\%$ and a specificity of 87.5% for enamel lesions and a sensitivity of 80% and a specificity $>99\%$ for dentine lesions were found using the NIR spectral imaging method.

Fig. 6

Correlation plot of the histological score and the average maximum NIR caries score, $S_{\text{caries}}$ , within the map region corresponding to the histological section.

4.

Discussion

There is great interest to detect early stages of tooth decay because it is reversible, and the progression to a stage where restorative intervention is required can therefore be avoided. Screening early lesions in occlusal pits and fissures is a main concern because these regions have a higher susceptibility to bacterial deposition and therefore demineralization. Radiography and visual inspection are commonly used in the clinic to perform this task; however, the former method lacks the sensitivity needed to detect early stages of the disease and both lack the ability to quantify its progression successfully. Although FOTI has been used to improve the detection of caries in occlusal and interproximal surfaces,^{9, 16} quantification is still an issue and the excess of light scattering and therefore reduced penetration in visible wavelengths $(400–700\mathrm{nm})$ imposes a performance limitation. Moreover, stain is a strong confounding factor that affects this technique.

The most promising quantitative methods for occlusal caries detection previously reported include QLF,²⁶ laser red fluorescence,²⁷ and ECM.¹⁰ Although OCT is still in an experimentally stage, it has also shown the ability to examine tooth structure non-invasively.^{25, 28} The downfall of electrical caries monitoring is that it is a single-point measurement and not much information about the lesion distribution can be obtained. Laser red fluorescence is also single-point technique and is mainly designed to detect bacterial invasion, but a strong fluorescence signal is normally obtained only when the lesion is already in the dentin. QLF is mainly designed for the detection of enamel lesions; however, stains represent a strong confounding factor for this technique. Bühler ²⁹ have recently introduced NIR imaging as a method for occlusal caries detection. This method operates in the same manner as FOTI but employs NIR wavelengths and therefore improves the light penetration through the enamel, increases the image contrast, and can reveal the presence of hidden lesions. The technique was therefore proposed by the authors for delineating the lesions and possibly used in combination with OCT. Although excellent results are reported with NIR imaging, quantification of the lesions is difficult when employing a transillumination configuration due to the shadows created within the samples and the image bleaching caused by the side illumination. An improved uniform illumination was however obtained by using a cylindrical collimating lens.²⁹ These results are nevertheless encouraging, but only wavelengths of 830,³⁰ 1310, and $1550\mathrm{nm}$ (Refs. 21, 29) have been explored in the NIR for caries detection.

Hyperspectral imaging is a powerful method used to interrogate the spectral characteristics of samples in a two-dimensional space. This feature is of particular use when studying teeth due to their inherent heterogeneous occlusal geometry and associated lesion distribution. We have employed this method for our investigations to demonstrate the ability of NIR spectral imaging to quantify caries lesions from occlusal surfaces. Note that we have used a reflectance configuration since uniform illumination of the tooth is easily achieved. Imaging with NIR wavelengths is of great interest since stain no longer represents a strong confounding factor when detecting tooth demineralization. Figure 7 shows an example of the reflectance obtained at a wavelength of $1250\mathrm{nm}$ for three heavily stained teeth. It is possible to observe that the absorption of light by stain is minimal and confirms previous reported observations employing NIR wavelengths.²⁹

Fig. 7

NIR images of highly stained occlusal tooth surfaces taken at $1250\mathrm{nm}$ . It can be clearly observed that the absorption by stain is minimal and shows little influence to reveal underlying tooth decay.

Note that the spectral intensity dips observed in the reflectance spectra shown in Fig. 3 correspond to the absorption peaks of water. In addition, the expected raise in light scattering caused by white spot lesions can be observed in the spectra as a background intensity across all wavelengths. The results obtained for the different lesions suggest that, as the cavity reaches the dentin, the lesion size increases and the amount of water within also increases. For early enamel demineralization, the effect is rather observed as an augmented light scattered intensity at the surface due to the porous structure of the lesion. These physical effects are used as a mechanism to quantify the extension of tooth decay; in particular we proposed to use the reflectance at 1440 and $1610\mathrm{nm}$ as presented in Eq. 4 since these wavelengths prove to be most affected by water absorption and scattering, as shown in Fig. 3, for enamel and dentin lesions. Sound regions of the teeth show a reduced reflectance at wavelengths above $1450\mathrm{nm}$ ; this might be caused by an increase in the absorption of light by hydroxyapatite and/or collagen; decayed areas have a reduced amount of mineral and/or organic material, and this can explain the observed higher reflectance at such wavelengths.

Although special attention was paid to remove the influence of specular reflections in our algorithm, the measurements were still affected by this source of noise; in particular, at the edges and crests of the tooth where strong reflections are expected. Further improvements on the setup are in progress, including the addition of polarizers. Note that $S_{\mathit{\text{caries}}}$ in Eq. 4 uses wavelengths below $1.7\mu \mathrm{m}$ . We have generated a second algorithm that performs equally well and has little influence by specular reflections; however, the algorithm requires wavelengths above $1.7\mu \mathrm{m}$ . A current drawback of NIR spectral imaging is the cost of the detector; CCD cameras do not work in this wavelength range, and therefore, InGaAs or MCT cameras are normally used. Note that InGaAs has a spectral response that covers up to $1700\mathrm{nm}$ , whereas MCT covers up to $2500\mathrm{nm}$ ; the later being two- to threefold more expensive than the former. We have therefore developed the algorithm that could be used with an InGaAs chip.

Note that the number of teeth used in this investigation was limited by the availability, and additional tooth samples could be sourced and used in the future for further validation studies. In addition, tooth sample storage methods should ensure that its organic component (mainly collagen) is preserved. Further studies are therefore required to examine the influence of different storing methods³¹ on NIR spectral imaging; however, we expect little influence since this is not a fluorescence technique where protein denaturalization could affect the signal significantly.

Although the method fails to score correctly where there is a tooth hole, this does not represent a major drawback since it can be used as an adjunct to visual inspection where this type of lesion can easily be detected.

We believe that the NIR spectral imaging method presented can be an important part of the diagnostic armament of dental practitioners and can potentially improve and simplify the decision-making process, leading to better treatment planning and to the reduction of unnecessary invasive procedures. Further in vitro studies to understand the relationship between this method and mineral loss by using microradiography are still needed. In addition, we believe that lesion progressions could also be followed with this method due to its quantitative nature, and this can be of great use for clinical trials; additional investigations are however required to support this capability.