The findings of this study support the potential use of fluorescence imaging to objectively quantify dental fluorosis. This is consistent with earlier work [15
]. However, the correlation coefficients in the current study are lower than those obtained by Pretty et al [15
]. This is probably due to the fact the population in the original study was a selected population based upon the presence of fluorosis in an area of optimal water fluoridation and presented with only milder forms of dental fluorosis. The population in the current study is larger and presents with a greater range of fluorosis severity and the increased presence of confounding factors. Nevertheless, the repeatability of both techniques is very good with the ICC for the existing technique being commensurate with the findings of Pretty et al and the convex hull software delivering even greater performance. This was achieved without employing techniques such as video repositioning and as such supports the claim that fluorescence image analysis can be robust in terms of the repeatability of measures [15
There are certain considerations to be made regarding the population selected in this study. The population was selected according to the level of fluoride in their cooking water. Despite the fact the TF score obtained from the photographs was able to separate the different water fluoride content intervals (Figure ) (suggestive of a dose response) it is clear this is not a true reflection of the fluoride exposure of the subjects. The risk to developing enamel fluorosis must include all forms of fluoride ingestion at the time of tooth development not only from cooking water but also drinking water, beverages, food and oral hygiene products [19
]. It would be problematic to use total fluoride exposure to assess dose response in this population on this study and it should be accepted the use of cooking water fluoride content is not indicative when evaluating a dose response. However, this population was selected as lifetime residents and the likelihood the cooking water source had changed since birth was low. It had also been demonstrated that the current cooking water fluoride content was a strong measure when determining fluorosis risk [22
Looking at the ranges of water fluoride content of the intervals (Table ) intervals 0 and 1 could be seen to represent non-fluoridated populations with perhaps some background fluoride in water. Intervals 2 and 3 are commensurate with sub-optimal and optimally fluoridated populations with interval 4 representing fluoride levels above optimal levels. It would be desirable that any system would be able to discriminate between each of the intervals. However, it could be argued at the levels set in this study the difference between intervals 0, 1, 2 and 3 is minimal and the inability to discriminate between intervals 0 and 1 is not critical. However, a robust system should be able to discriminate between interval 4 and the remaining intervals.
Whilst the outcome of this study supports the development of fluorescence imaging as a technique for objectively quantifying enamel fluorosis, there remain several unresolved issues from the work of Pretty et al. Firstly there is still no acceptable gold standard to use. The use of the photographic score as the comparator remains inadequate as it depends upon a subjective assessment of fluorosis. The conventional digital photograph requires the camera to be position at an angle to the teeth (approximately 15° to the perpendicular plane) to reduce specula reflection, whereas the fluorescence imaging uses flat field illumination and polarizing filters enabling the images to be captured perpendicular to the teeth. This results in potential differences in the information that can be displayed between the images owing to foreshortening of the photographic image. Furthermore it is still not possible to relate the TF score from the photographs to the metrics obtained from either of the fluorescence analysis techniques. This is not a situation unique to the assessment of fluorosis, similar issues existed when fluorescence imaging was used for the assessment of carious lesions [23
]. This would be true of any novel technique utilizing emerging technologies. Nevertheless, both fluorescence imaging techniques demonstrate an increase in ΔQ with increasing TF index score (Figures
The decision to base the analysis on remote consensus scoring of standardized photographs was justifiable owing to the reduction of bias and examiner thresholding. The quality of the photographic images still enabled the detection of focal loss of surface enamel. Any issues associated with potential loss of validity were addressed by grouping the data for subjects with a TF score or 4 or higher. There is additional justification for this owing to the age range of subjects and the fact TF scores greater than 4 generally present as a result of post-eruptive changes to the fluorotic enamel.
The statistical analysis of the data is also compromised by the differences in the metric outputs. The correlation coefficients presented in this paper should not be regarded as a measurement of agreement as they are merely an indication of association between the different techniques. This is not only true of the comparison between the photographic scores and the fluorescence imaging but also between the two fluorescence imaging techniques. Despite similarities between the fluorescence imaging techniques, the methods by which the metrics are derived differ. The outputs whilst delivering the same outcome measures are presented using different scales.
All of the above factors contribute to difficulties in assessing the sensitivity and specificity of the fluorescent imaging technique and software analyses. In order to estimate the sensitivity and specificity of the fluorescent imaging system the data was exported to Stata (release 11, StataCorp, TX USA) and ROC curves produced using classification models for the QLF metric output ΔQ for each software analysis technique and a classifier boundary, or threshold, for fluorosis (TF score) of ≤2 and ≥3. These ROC curves are illustrated in Figures
and . The data would suggest (from this rudimentary assessment of sensitivity and specificity) the convex hull software demonstrated higher levels of sensitivity and specificity (sensitivity 80.61%; specificity 80.96%) when compared to the existing technique (sensitivity 68.37%; specificity 83.27%). The outcome was similar for estimating accuracy when comparing the area under the curve (AUC) for the convex hull and existing technique (0.8802 and 0.8086 respectively).
ROC curve for convex hull software.
ROC curve for existing technique software.
In order to reduce variance between the two fluorescence imaging techniques it was necessary to utilize the same masks of the teeth. The software in the original study required the operator to draw around the object teeth with a region interest tool. It is clear that repetition of this process could result in variance. Furthermore the original software required a reference area to be selected using the region of interest tool. This was overcome by using software written in Visual C# (2005 Express Edition, Microsoft, Inc., CA, USA) to process masks for all the object teeth from the fluorescence images. The software for the existing technique was augmented by the addition of an algorithm written in MATLAB that automatically selected a reference area from the triangulation of a point located on the gingival tissues with the masks of the maxillary central incisors (with an assumption of the location of the teeth). This algorithm worked well but was unable to process the analysis if there was either a missing mask (missing, fractured or restored incisor) or there was a large diastema between the central incisors. If this occurred the subjects and data were excluded from the analysis. This resulted in the exclusion of seven subjects.
The inability of the fluorescence imaging techniques to differentiate fluorosis from caries and other non-fluorotic developmental defects of enamel still exists. The subjects illustrated in Figure demonstrate issues that can arise from this phenomenon. The images of subject 545 illustrate how the presence of caries and stain can impact upon the fluorescence image and subsequent analysis. The presence of plaque, stain, caries and other developmental defects of enamel such as demarcated enamel opacities are confounding factors in fluorosis assessment using fluorescence imaging [24
]. It has been shown that demarcated opacities with similar clinical presentations can exhibit markedly different changes in fluorescence with some opacities demonstrating a loss of fluorescence whilst others demonstrating an increase in fluorescence signal.
Figure 7 Images of subjects with confounding factors for QLF. a. clinical photograph subject 837 presenting with non-fluorotic hypomineralization and enamel loss on maxillary right central incisor. b. QLF image of subject 837. Note the pattern of fluorescence (more ...)
Subject 837 (Figure ) had suffered from a large developmental defect localized to the right maxillary central incisor with an aetiology non-fluorotic in nature. Both imaging techniques were unable to differentiate this from fluorosis and hence large values for Area, ΔF and ΔQ whereas the score allocated from the photograph for this subject was TF 0.
The images of subject 230 (Figure ) illustrate fluorosis that has developed post eruptive stain. Whilst the existing technique was able to process this image the convex hull software was unable to differentiate the change in fluorescence relative to the surrounding unstained fluorosis and would have deemed the areas of discolouration as “heavy stain” and allocated a higher score for ΔF and ΔQ accordingly.
It is clear further work is needed if fluorescence imaging techniques are to be used for objectively quantifying fluorosis. It has been shown it can discriminate between populations with differing fluoride exposures. It is arguable which analysis technique is the more appropriate technique. The convex hull software would appear to be more sensitive than the existing technique at low fluoride exposures. This is likely to have been caused by the low threshold level set on this study. This was necessary to avoid excluding milder forms of fluorosis but would have included greater levels of noise in the analysis, affecting specificity. In fact the data suggests the ability of the convex hull software to discriminate between levels of fluorosis severity is comparable to the use of photographic scores. The existing technique appears to work well at higher severities of fluorosis. This is in contrast to the findings of Pretty et al who hypothesized that artifacts created by the existing technique may underestimate fluorosis. This may have been based on the findings from a population with lower exposures to fluoride and lower severities of fluorosis presentation. Overall both fluorescence image analysis techniques appear to be less sensitive than clinical judgment using an index when considering the whole range of presentations of fluorosis. Although in the case of the convex hull software this is marginal.
Although image capture is simple and reproducible it remains an additional step in study procedures. In addition, despite the fact the analysis is automated, there remains a considerable operator task in drawing the masks for image processing. At present it would appear the use of at least a photographic score using TF index and the application of diagnostic criteria cannot be dispensed with. The question arises as to what additional value can the use of fluorescence imaging provide over and above a clinical index? The answer may lie in the fact the longitudinal assessment of fluorosis is desirable and the variation in examiner scoring using a clinical index could be problematic when assessing prevalence and severity by clinical examination [9
]. This can be avoided with the use of photographic scores, but the problem of subjectivity would remain.
Further software development is required particularly with respect to the production of the masks of the object teeth as this is the time dependant process that questions the viability of the application in a large epidemiological survey. Possible avenues to explore would be the production of automatic masks using edge detection software or more simply the use of preset polygons in Visual C# that can be adjusted to the shape of an object tooth rather than masks drawn freehand.
A possible interim solution could be to use a dual-camera system for image capture using two high resolution CCD cameras with an illumination and lens array that would permit one camera to capture a fluorescence image and a second to capture a polarized white light image (negating the need for camera repositioning to reduce specula reflection). Both sets of images would be of the same position relative to the teeth, same magnification and would both be amenable to longitudinal assessment through the use of video-repositioning software. Any white light image score using an index can remain blind and randomized and quantifiable metrics of fluorosis obtained from the corresponding fluorescence image.