1. Introduction

Dental caries is a major oral health problem experienced by over 90% adults aged 20-64, 36% of children aged 2-8 (on primary teeth), and 21% of children aged 6-11 (on permanent teeth) in the USA [1,2]. This high caries prevalence strongly suggests that the current caries prevention and management is unsatisfactory and more advanced strategies are in need. The most effective way to manage dental caries is to detect them at an early stage when they can be stopped or reversed through noninvasive treatments by fluoride, ozone, or minimal intervention (MI) paste [3–6]. Usually, the earliest caries that can be detected via visual and tactile inspection are chalky white lesions, which are difficult to be identified and thus often overlooked by dentists during the regular dental visits [7]. By the time when these lesions are caught by naked eyes, dental probing, or radioactive imaging (e.g., x-rays) later, they typically require invasive fillings, which may fail over time and lead to more advanced restoration methods such as crown, bridge and dental implant therapy [8,9].

Tremendous efforts have been invested in the development of optics based spectroscopy and imaging technologies for early caries detection because of their remarkable benefits in non-invasive and non-contaminant measurements and real-time analysis [10–14]. Fluorescence based technology is currently the most popular optical approach for caries detection, which is evidenced by commercial products such as KaVo DIAGNOdent, Spectra from Air Techniques, and Caries I.D [14–17]. The fluorescence-based device is highly sensitive for caries because the fluorescence level of enamel increases as caries develop. However, these devices can be easily disturbed by the fluorescence originated from other oral conditions such as prophy paste, stained fissures, calculus and plaque etc., which often lead to false positive diagnosis [18].

Raman spectroscopy is a promising technique for early dental caries detection as it can detect the change of mineral amount in enamel which is caused by early caries. Enamel is the highest mineralized tissue in human body, and at maturity it contains 96% inorganic mineral by weight (more than 86% by volume), which primarily exists in the form of Ca₁₀(PO4)₆(OH)₂, better known as hydroxyapatite. Enamel also contains small portion of organic matrix and 4 - 12% water by volume, which exists in the intercrystalline spaces among a network of micropores [19]. Raman spectroscopy can be effectively used to characterize PO₄^3- vibrations in the mineral phase of hydroxyapatite and detect the changes of their intensities, which can serve as the indication of early carious lesion [20]. A previous study determined that caries analysis using polarization resolved Raman spectroscopy yielded > 95% sensitivity and > 95% specificity [10]. However, traditional fiber based Raman probes can only focus the laser beam to tiny spots (typically << 1 mm in diameter), while the typical size of occlusal surface of human molar is 7-9 mm in diameter [21]. In order to have a large-scale inspection over the tooth surface, the traditional Raman scanning method is extremely time-consuming, and the Raman probe can be easily misplaced due to the small size of early caries. Thus, expeditious imaging of the tooth surface is the major roadblock preventing the translation of Raman spectroscopy to the clinical application. Yang et. al. demonstrated that wide-field Raman imaging is orders faster than point-scan method, however, the laser could still only illuminate the central area that is ~1% of the full field of view due to the converging effect of the imaging lens and the limited laser power [22]. This work demonstrates a hyperspectral Raman imaging system after minor, but important, modifications from our previous report, which led to significant advancement in the performance of Raman imaging: it allows obtaining Raman images over the entire tooth surface within seconds, making it feasible and convenient for caries diagnosis at dental clinics.

2. Methods and material

Raman images were constructed based on the integrated Raman signal relating to the phosphate symmetric stretch bond of mineral, which appears at ~960 cm⁻¹ in the Raman spectrum. The Raman images were acquired via an anti-reflection achromatic lens (f = 150 mm), whose long focal distance allowed placing the dichroic mirror between the lens and the sample. Thus, the illumination area was the same size as the laser beam diameter, which had been adjusted to be approximately 1 cm in diameter on the sample surface. A high power tunable Ti:Sapphire laser (Spectra Physics, 3900S) operating at 785 nm was used as the excitation source and an EMCCD (iXon 888, Andor) having both EM amplified mode and conventional mode was used to capture Raman images (Fig. 1(a)). An edge filter at 785 nm was inserted to reject the Rayleigh scattering from the sample, and a narrow band pass filter was used to pick up the 960 cm⁻¹ mineral Raman signal associated with the mineral phase (Fig. 1(b), blue curve). The second acquisition at the valley of the mineral peak at ~880 cm⁻¹ (by slightly rotating the narrow band pass filter) was conducted to obtain the background fluorescence, see Fig. 1(b), red curve. Subtraction of the two acquisitions resulted in a 2D image displaying the distribution of the mineral intensity. Integration time varied from 1 to 30 seconds in this study to demonstrate the capability of the system for fast Raman imaging or the acquisition of high quality Raman images. The Raman signal was routed to a Raman spectrometer (SuperGamut 785, Bayspec) to verify the complete removal of the Rayleigh scattering and the correct selection of the Raman peaks before image acquisition. The laser power was set in the range of 600 - 800 mW (corresponding to 0.76-1.0 W/cm² laser intensity on the sample) in this study so that the Rayleigh scatter (i.e., residues of the laser light) was completely blocked by the Raman filters. The iXon 888 camera has 1024 × 1024 pixels, and the pixel size is 13 µm. Thus, the full-scale imaging size (6.66 mm × 6.66 mm) was half of the chip size when an f = 300 mm lens was used to focus the images onto the 2D camera.

 

Fig. 1 (a) Schematic layout of the imaging setup. (b) Typical Raman spectra of teeth when the band-pass filter is adjusted to center at 960 cm⁻¹ (blue curve), 880 cm⁻¹ (red curve) and removed (green curve).

Download Full Size | PDF

A buffalo tooth was embedded and stabilized with epoxy resin and sliced to expose the dentin-enamel junction (Fig. 2(a)) which served as the standard sample demonstrating the sharp transition from high level mineralization (enamel) to low level mineralization (dentin). Two human teeth were collected from a local oral surgeon clinic under the IRB approval of Jackson State University. One tooth had a clear white lesion, and the other tooth appeared healthy and solid. The teeth were imaged to demonstrate the capability of the Raman system to image the full-scale tooth surface.

 

Fig. 2 (a) Picture of the sliced buffalo tooth under white/flash light illumination. (b) Representative Raman spectra from dentin and enamel side of the tooth. Full-scale Raman images of the buffalo tooth when the band pass filter was tuned to (c) signal peak (I_sig) at 960 cm⁻¹, (d) fluorescence reference (I_ref) at 880 cm⁻¹, and (e) their difference (I_sig-I_ref). (f) A laser intensity distribution (I_laser) image was acquired to obtain (g) the laser intensity calibrated ((I_sig-I_ref)/I_laser) mineral distribution image, which is then normalized and re-plotted in (h) 3D format.

Download Full Size | PDF

3. Results

The sliced buffalo tooth was firstly used to validate the integrated Raman system, and an area across the boundary between the enamel layer and the dentin region (dashed red box in Fig. 2(a)) was selected for acquiring Raman images. Representative spectra of enamel and dentin were provided (Fig. 2(b)). Both enamel and dentin were not uniform in terms of Raman intensity and fluorescence background level, thus spectra from 10 random points were taken from each side, and the demonstrated spectra were the averages of each set. The spectrum from each point was taken with 30 mW laser illumination through a 10x objective lens, and the integration time was 1 second with 9 times of average. The power of the illumination laser light for acquiring full-scale Raman images was 600 mW on the sample surface, and the integration time of the camera was set at 30 seconds. Full-scale Raman images were taken from the selected area with band pass filter tuned to be at 960 cm⁻¹, the Raman signal peak (Fig. 2(c)), and 880 cm⁻¹ (Fig. 2(d)), the fluorescence baseline (as reference), respectively. The difference of the two images was the fluorescence background subtracted image reflecting the mineral distribution (Fig. 2(e)). Figures 2(c), 2(d), were normalized in terms of the maximum value of Fig. 2(c). Figure 2(e) was shown in a different range to present appropriate contrast. Some places on the dentin side appeared brighter than the enamel side (Fig. 2(c)) because the dentin was more fluorescent than enamel as shown in background fluorescence image (Fig. 2(d)). The central area of the fluorescence background subtracted image (Fig. 2(e)) was brighter than the peripheral area, suggesting the nonuniform distribution of the laser intensity over the field view. Considering Raman scattering is linearly dependent on the incident laser intensity, an image of laser intensity distribution (Fig. 2(f)) was acquired with Raman filters (i.e., the edge and band pass filters) replaced by a OD8 neutral density filter. The exposure time was decreased to the level where the CCD camera was far below saturation. The enamel portion appeared darker than the dentin portion in the laser intensity distribution image indicating less photons were backscattered from the enamel. This observation was consistent with the smaller scattering coefficient of enamel discovered by Daniel Fried etc [23]. The laser intensity distribution was used to calibrate the fluorescence subtracted Raman image by division, as shown in Fig. 2(g). This step corrected the non-uniformity of laser distribution due to optical distortion and spatial variation on sample surface, which enhanced the contrast between enamel and dentin. At last, the laser intensity calibrated mineral distribution image was normalized, blurred with Gaussian filter, and replotted in the colored 3D format for additional viewpoints (Fig. 2(h)).

The same imaging procedure applied to the buffalo tooth was repeated on a human tooth- an incisor, which had developed an over 1 mm white lesion (see Fig. 3(a)) on one of its side surface. The power of the illumination laser light was kept the same at 600 mW on the sample surface, and the integration time of the camera was also set at 30 seconds. The white lesion visible to naked eyes was shown in the picture taken by a cell phone camera (Fig. 3(a)).

 

Fig. 3 (a) Picture of the incisor with white lesion taken by cell phone camera. Full-scale images covering the lesion when the band pass filter was tuned to (b) Raman signal peak, and (c) fluorescence reference. Fluorescence subtracted images (d) before and (e) after laser intensity correction. The laser intensity corrected Raman image was replotted in (f) in 3D format for additional viewpoints.

Download Full Size | PDF

Figures 3(b) and 3(c) presented the full-scale images when the tunable filter was respectively set for the Raman signal and fluorescence reference correspondingly, and they were normalized by the maximum of Fig. 3(b). The lesion was confirmed by fluorescence subtracted Raman image (Fig. 3(d)), which was not normalized to the maximum of Fig. 3(b) because the contrast was too low. The laser intensity corrected Raman image demonstrated the improved contrast between normal and caries regions (Fig. 3(e)). The normalized 3D format reprint (Fig. 3(f)) of the laser intensity corrected Raman image suggested that the tip region of the tooth had significantly less mineral concentration, which was hardly seen on the 2D image.

The second tooth studied with the system under the same procedure was a premolar. The experimental conditions were modified: the laser power was increased to be 800 mW while the integration time was set at 10 seconds. This second tooth did not have a noticable lesion like the above incisor, and it would more likely be diagnosed as a healthy tooth under visual examination (Fig. 4(a)). The Raman images, before (Fig. 4(b)) and after (Fig. 4(c)) laser intensity correction, indicated that the mineral intensity in the marked region (dashed orange lines) was significantly lower than the rest of the tooth, suggesting a caries was probably forming in this area. In the meantime, we noticed cusp tips also showed lower mineral intensity (marked with red circles, Fig. 4) than surroundings, which was possibly due to increased occlusal tooth wear and possible dentin exposure. The small region on the top right corner had the highest mineral intensity, but the reason is unclear at this moment.

 

Fig. 4 (a) Picture of the premolar taken by cell phone camera. Full-scale fluorescence subtracted Raman images of the top surface (b) before and (c) after laser intensity correction. The laser intensity corrected Raman image was replotted in 3D format (d) for additional view point.

Download Full Size | PDF

4. Discussion

In this work, we preliminarily verified the feasibility of dental caries detection by using full-scale Raman imaging and reported the results from two representative tooth samples. In addition to high sensitivity and high specificity, the full-scale Raman imaging has a strength of inspecting an area of interest within a short amount of time at a high resolution, unlike the point by point inspection of traditional Raman spectroscopy. Benefit from the full field of view illumination, the resolution of full-scale Raman imaging is mainly constrained by the diffraction limit of the objective lens and sub-micron resolution can be achieved. In macroscopic full-scale Raman imaging as in this work, the resolution limitation of the digital CCD camera overtakes the limitation of the light diffraction. However, lesions as small as 20 µm still can be resolvable, since each pixel on the image corresponded to a ~7 µm region on sample surface. Usually, lesions smaller than 500 µm are ‘invisible’ to the visual and tactile examination [24], which are also very likely missed by point scanning based optical techniques. Resistance to the interference from water in oral condition is another strength of Raman imaging. Some lesions only appear under dry condition, while others are difficult to be recognized under wet condition [24]. Further, the specular reflection of light on the wet surface of tooth disturbs the visual examination, but it has no influence on Raman imaging measurement.

In direct Raman imaging system, the images are constructed with the number of photons composing of the selected Raman signal. Effectively rejecting photons from other sources is critical because the imaging camera cannot discriminate Raman photons from others with different origins. Sources of unwanted photons include residues of excitation laser, ambient lights, unselected Raman signals, and laser induced fluorescence. Among these sources, laser residuals and unselected Raman signals were completely removed by the narrow band pass filters, as confirmed by the Raman spectra after the filtration. The ambient light was negligible in dark room after tube and cage sealing. The laser induced fluorescence was composed of two parts, i.e., photons from selected region and photons from rejected regions of the band pass filter, respectively. Although the Raman spectra after narrow band pass filter (dashed blue and red curve in Fig. 1(b)) demonstrated almost zero level of the fluorescence baseline out of the selected region, we speculated a significant number of photons were collected from the rejected regions. This speculation could explain the discrepancy between the Raman spectra and Raman imaging. The Raman spectra predicted comparable photon counts between Raman image (i.e., I_sig-I_ref) and its corresponding fluorescent opponent (I_ref), but the intensity in their corresponding full-scale images presented up to an order difference. However, the leak of the fluorescent light would not negatively impact our concept of Raman imaging because these photons were removed from the image after fluorescence subtraction. Therefore, the fluorescence subtracted Raman image remained an authentic presentation of mineral distribution.

Accompanying the fluorescence was the influence of photobleaching effect, which might potentially alter the interpretation of the data. This effect usually becomes problematic at long integration time and/or high laser power, and would appear more serious on caries (or dentin) than normal enamel. For short integration time (e.g., 1 second), the photobleaching effect was negligible on enamel with or without small white lesions. As the integration time significantly increased (e.g., to 30 seconds), a waiting time could be applied before imaging acquisition until the change of the fluorescence baseline was not measurable in the Raman spectrum. For teeth samples with large lesions or discolored cavities, this waiting time was up to 5 minutes. Excitation by lasers at longer wavelength, e.g., 1064 nm would be beneficial for fluorescence reduction and waiting time for photobleaching would not be needed [25]. However, the Raman scattering efficiency would be reduced under 1064 nm excitation, and the increased readout noise and sensitivity to ambient light could further affect the interested Raman signals. Other wavelength longer than 785nm may be considered, but the only consideration for those choices is the tradeoff between the reduced fluorescence and the reduced sensitivity of CCD detectors at longer wavelengths.

Total time required to investigate a tooth is a critical consideration for clinical practices, because keeping the tooth immobile for diagnosis over extended time will place extra burden on patients. The total imaging times for the above two representative human teeth were 60s and 20s (i.e., the addition of the time for I_sig and I_ref images) in order to present high quality images with decent contrast between decayed and normal enamel. The shortest integration time tested in this study was 1s (on buffalo tooth, Fig. 5(a)), which yielded lower signal to noise ratio as expected. In an effort to improve the signal collection, we investigated the performance of the iXon camera in EMCCD mode. However, the comparison between the conventional CCD mode and EMCCD mode suggested that EM amplification was not beneficial in the acquisition of the full-scale Raman images. As shown in Fig. 5, the image of Fig. 5(a) was acquired under CCD mode operation without amplification, while the image of Fig. 5(b) was taken under EMCCD mode operation with × 100 amplification, and the two images appeared to have the same level of signal to noise ratio. We tested various integration time on human teeth as well. For example, Fig. 5(c) shows the laser intensity corrected image of the decayed incisor taken with 5s integration. This integration time will lead to ~10s for acquiring one complete analysis for one surface, and we would assume this is practical in clinics.

 

Fig. 5 Full-scale Raman images of sliced buffalo tooth when camera was operated in (a) conventional CCD mode and (b) EMCCD mode. The integration time was 1s and the image was calibrated with the same laser intensity distribution shown in Fig. 2(e). (c) A full-scale Raman image of decayed incisor taken with 5s integration.

Download Full Size | PDF

We note that the maximum laser intensity employed was ~1.0 W/cm², which exceeds the maximum permissible exposure (MPE) of skin by a factor of ~2 (ANSI Z136.1-2014). However, this intensity is well below the damage thresh hold of skin, and it is approximately three orders lower than the intensity for orthodontics applications [25–27].

5. Summary

We demonstrated that full-scale Raman imaging covering the cross-section area of the laser beam can be achieved by employing a high-power laser, which is guided to the sample without traveling through the objective lens. The full-scale Raman images based on the mineral intensity distribution can be used to separate the enamel from dentin, as well as normal enamel from decayed enamel. More importantly, the Raman images may identify dental caries that are not visually detectable. In summary, this full-scale Raman imaging is a promising technique for caries detection at dental clinics.