Autofluorescence in the Oral Cavity
Introduction
In recent years, dentists have begun using the VELscope Vx, a handheld scope that enables fluorescence imaging to visualize oral mucosal abnormalities, in oral exams. This technology enables imaging of abnormalities that may not be visible to the naked eye and has enhanced diagnosis of oral cancer and dysplasia. This device offers clear proof that there is interesting health data contained in oral fluorescence, and this project seeks to pursue that information. Specifically, this project aims to determine what data exists in the fluorescent spectrum, how we can reliably obtain this data, and how this can be used in diagnosis of oral disease.
Background
In 2003, Veld et.al. conducted a comprehensive study aimed at characterizing the autofluorescence of healthy oral mucosa in different parts of the mouth. This group recognized the power of autofluorescence spectroscopy in oral cancer detection and hoped to construct a reference database of fluorescent spectra from healthy oral mucosa. In this study, spectra were recorded from 97 volunteers under a selection of excitation wavelengths between 350 and 450 nm at different anatomical sites. Of particular interest to this project, the fluorescence measured on the dorsal side of the tongue exhibited a broad peak between 500 and 550 nm and a sharp peak above 600 nm.
In 2006, Wu and Qu studied autofluorescence of human epithelial tissue and obtained a similar broad fluorescent spectrum around 500 nm in oral tissue excited at 405 nm. They hypothesized that this peak may be due to overlapping of NADH, FAD, and keratin fluorescent spectra. They also observed a sharp peak in esophageal fluorescence around 680 nm and hypothesized that this peak may be attributed to porphyrin derivatives.
In 2017, Bjurshammar published a thesis exploring “Porphyrins and Phototherapy of Oral Bacteria.” This study discussed the spectral properties of porphyrins, a group of macromolecules found in most bacteria typically in the oral cavity. Notably, protoporphyrin IX has a sharp fluorescence emission peak around 635 nm and is efficiently excited at 405 nm. The data from these studies is shown in Figure 1. The results of all of these studies suggest to us that we can expect to see significant autofluorescence in healthy oral tissue, due either to autofluorescence of oral mucosa or to the presence of porphyrins. This project seeks to collect further data about fluorescence and its ties to oral health. Fluorescence on volunteers’ tongues was measured under 400 nm light, chosen because it will excite relevant fluorescent species and is visible, preventing damage from unseen light. This data was analyzed using singular-value decomposition and least squares estimation in an effort to elucidate connections between fluorescence data and oral health.
Methods
Data Collection
An experimental setup was designed to allow radiance data to be collected from a calibration spot and subject’s tongue under both blue and tungsten illuminants, with minimal time and effort required from the subject (Figure 2). Spectral data was collected using a SpectraScan 670 spectrophotometer controlled with a Matlab script. Subjects were given protective goggles designed to filter out blue light and sat at the edge of an optical table behind a chin rest. A 400nm blue light was mounted to the left of the chinrest, and a tungsten light (ARRI 1000 Plus) was placed on the far side of the optical table. The spectrophotometer was to the right of the subject.
Subjects were first asked to hold a Labsphere Reflectance Standard in the approximate location their tongues would be when using the chin rest. The spectrophotometer was focused onto the reflectance standard, and calibration reflectance curves were obtained under blue illumination, tungsten illumination, and combined illumination. The subject was then asked to place their chin in the chin rest and hold their tongue out. The spectrophotometer was focused onto a spot in the middle of the tongue, and radiance was again measured under the three illumination conditions.
Fluorescence Estimation
The simplest way to measure fluorescence would be to illuminate the tongue with a single wavelength of blue light and measure the radiance seen by the spectrophotometer. Any light in the visible spectrum (outside of the incident wavelength) could be assumed to be due to fluorescence. However, the blue excitation light we used had energy throughout the visible spectrum (Figure 3), and reflectance of this light needed to be accounted for. Tongue reflectance was therefore calculated using reference data taken under tungsten illumination. The reflectance of the tongue was calculated by dividing the spectrum of the tongue under tungsten light by the spectrum of the tungsten light reflected off of the reflectance standard. The predicted reflected light from our blue excitation light can be calculated by scaling the blue excitation light spectrum by each subject’s tongue reflectance function. Any radiance greater than this predicted reflectance can be assumed to be due to fluorescence.
Singular Value Decomposition
From a collection of fluorescence data, we can attempt to find the dimensionality of the data using singular value decomposition (SVD). Briefly, SVD can be used to find a set of basis functions, linear combinations of which can be used to approximate the collected data. In this project, finding a small set of basis functions to explain collected data may allow for the correlation of fluorescence data with specific fluorescent populations in the mouth (e.g. specific fluorophores, bacteria, etc.) and correlation to oral health diagnoses. SVD was performed in Matlab, and data was reconstructed from basis functions as a check.
Least Squares Estimation
SVD imposes few restrictions on the basis functions produced. The formulation of the decomposition is likely to suggest the highest singular value basis as simply the mean of the data. Choosing a set of basis functions and performing least squares estimation to find the associated coefficients will circumvent this. We therefore sought to choose a set of basis functions from previously collected fluorescence data and fit our data to these functions, allowing us to link data directly to the underlying biology. Least squares estimation was also performed in Matlab.
Results
Average Fluorescence
Data was collected from 13 different subjects, with one subject volunteering for data collection before and after brushing their teeth, for a total of 14 data sets. The fluorescence estimation was performed on all data sets, and an average fluorescence estimate was calculated (Figures 4 and 5).
Verifying Fluorescence Estimate
Because the fluorescence data was estimated, we sought to verify the estimate. First, we checked whether the radiance measured under both blue and tungsten light agreed well with the predicted radiance when we sum spectra of the individual illuminants scaled by the tongue’s reflectance (Figure XX). We observed good agreement between this prediction and measurement, with disagreements attributed to fluorescence. In order to check that the found fluorescence values were reasonable, we compared the average fluorescence data to fluorescence measurements previously taken by Veld et. al. (Figure XX). The shape of the fluorescence estimate matched remarkably well to this data, suggesting that our estimate is reasonable.
Singular Value Decomposition
Wu and Qu have previously suggested that the broad shorter wavelength peak between 500 and 550 nm may be due to overlapping NADH FAD, and keratin fluorescence, while the sharp peak above 600 nm may be due to porphyrin derivatives. Given this suggestion, it seemed natural to first limit our SVD to two basis functions. The two basis functions with the largest singular values were kept and used in reconstruction of the measured fluorescence approximation (Figure XX). The success of this reconstruction is likely due to the fact that the first basis function largely follows the shape of the mean.
While this set of basis functions worked well for many of the data sets, it was clear that certain data sets were much better explained by more basis functions. For example, Subject 005 was much better reconstructed by 4 basis functions than 2 basis functions (Figure XX).
This illustrated to us two fundamental approaches to choosing the number of basis functions. The first approach is to let the biology inform our choice of basis size - given that we expect n fluorescent species to be present, we expect n basis functions will explain our data. The second approach is to choose how well we want to explain the data and let the singular values dictate our choice of basis size. We can plot the cumulative variability described against the number of basis functions used, set a threshold for the variability we want to explain, and choose the lowest number of basis functions that will reach this threshold (Figure XX).
Least Squares Estimation
As we hope to use this information for diagnostic purposes, it would presumably be advantageous for our basis functions to reflect the biology. Therefore, we choose basis functions based on previously measured fluorescence data and performed least-squares estimation to find how well these functions could explain our collected data. This approach has the additional advantage over SVD that it eliminates the possibility that a basis function can simply follow the mean.
For this estimation, we chose to use the basis set of Wu and Qu’s reported fluorescence curve, representing the overlapping fluorecence of NADH FAD, and keratin, and Bjurshammar’s porphyrin emission curve. Coefficients were calculated using least squares estimation in Matlab, and the fluorescence data was reconstructed using these coefficients. The original fluorescence estimate data was plotted against reconstructions making use of SVD and least-squares basis sets (Figure XX). Qualitatively, there is reasonable agreement for both reconstructions, suggesting that choosing a basis set based on the biology is worth further consideration. To quantify the performance of each reconstruction, we calculated the Euclidean norm of the difference between our original data and our reconstruction. The Euclidean norm from the least squares reconstruction was found to be ~1.09 times the Euclidean norm from the SVD reconstruction with two basis functions. Since SVD finds two basis functions to minimize this norm, this close match in performance again suggests that our choice of basis set based on biology is very promising.