TongueSim: Development of an Automated Method for Rapid Assessment of Fungiform Papillae Density for Taste Research

Abstract

Taste buds are found on the tongue in 3 types of structures: the fungiform papillae, the foliate papillae, and the circumvallate papillae. Of these, the fungiform papillae (FP) are present in the greatest numbers on the tongue, and are thought to be correlated to the overall number of taste buds. For this reason, FP density on the tongue is often used to infer taste function, although this has been controversial. Historically, videomicroscopy techniques were used to assess FP. More recently, advances in digital still photography and in software have allowed the development of rapid methods for obtaining high quality images in situ. However, these can be subject to inter-researcher variation in FP identification, and are somewhat limited in the parameters that can be measured. Here, we describe the development of a novel, automated method to count the FP, using the TongueSim suite of software. Advantages include the reduction in time required for image analysis, elimination of researcher bias, and the added potential to measure characteristics such as the degree of roundness of each papilla. We envisage that such software has a wide variety of novel research applications.

Introduction

Taste buds are situated on the tongue and contain a set of taste receptor cells bunched together. The taste receptor cells project microvilli into a pore at the top of the bud, which allow the cells to come into contact with tastants present in the saliva ( Segovia et al. 2002 ). Taste buds are found in the tongue in 3 types of structures (papillae) on the tongue: the fungiform papillae (FP), the foliate papillae, and circumvallate papillae. FP, so-called for their mushroom-like appearance when viewed in transection, are thought to be correlated to the overall taste bud number. For this reason, the density of FP on the tongue is often used to infer taste function. This association is still unclear, because some reports ( Zuniga et al. 1993 ; Bartoshuk et al. 1994 ; Doty et al. 2001 ), but not others ( Garneau and Derr 2013 ; Feeney and Hayes 2014a , 2014b ), show a correlation between taste intensity and FP density.

Historically, videomicroscopy techniques were used to assess FP density in a clinical setting. However, this method has several drawbacks, being time consuming (taking up to 30min to collect images) ( Shahbake et al. 2005 ) and difficult to conduct in children ( Shahbake et al. 2005 ). Videomicroscopy is also difficult to apply to research outside of the clinical setting, as it requires bulky equipment. More recently, advances in digital still photography and in software have allowed the development of rapid methods for obtaining high quality images in situ ( Shahbake et al. 2005 ). This technology has been used to measure taste papillae numbers in a wide variety of age groups ( Shahbake et al. 2005 ; Correa et al. 2013 ; Feeney and Hayes 2014a ; Feeney et al. 2014 ; Garneau et al. 2014 ). Although such methods are faster than the previously available techniques, they remain somewhat time-consuming, and are still subject to issues such as researcher bias, because the identification of FP can be subjective. These methods are also limited in terms of the various parameters that can be measured. For example, the shape of the papilla or the “degree of roundness” has previously been suggested as potentially having an effect on taste function, with children having been observed to have rounder papillae ( Segovia et al. 2002 ). To date, this hypothesis has not been fully explored. Indeed this factor could potentially explain the different results between some of the studies described above in the association between FP density and taste perception. Here, we describe the development of a novel, automated method to assess the density of FP, using TongueSim (freely available for download for research purposes, http://www.tonguesim.thinkbiosolution.com ). Advantages of this method include not only the elimination of researcher bias and a dramatic reduction in the time required to process the images, but with the added ability to measure the degree of roundness of each papilla, among other benefits. The software described here may provide an opportunity to study these, and other hypotheses that were not previously feasible with manual counting methods.

Materials and methods

Acquisition of tongue images

The images for this analysis were obtained from a previous study examining the relation between taste perception and FP density ( Feeney and Hayes 2014a ). All procedures were approved by the Pennsylvania State University Institutional Review Board (protocol number 33176). Participants were invited to take part via email, and written, informed consent was obtained from each. The front section of participants’ tongues was stained, using blue coloring on a sterile cotton swab, which allowed visualization of the FP. The participants held their tongue between a specially constructed transparent holder, with a 6mm diameter white adhesive circle attached as a size guide ( Feeney and Hayes 2014a ). Keeping the tongue held in the transparent holder, the participant rested their chin on a table-mounted chin rest. About 3–5 photographs were taken of the tongue in a flat plane, using a Canon EOS Rebel T3i camera, in manual mode with the following settings: ISO 800, Aperture: 29 (F29), shutter speed: 1/200.

Manual selection of FP

To identify the coordinates of the FP on the images, FP were first manually selected from 9 tongue images. This was conducted using the TongueSim graphical user interface (GUI), shown in Figure 1 . The resultant images, depicting the selected FP, are shown in Figure 2 . The manually counted numbers of FP for each set were used as standards to compare with results of automatic selection routines. All codes were written using a commercially available suite of programme (MATLAB and Statistics Toolbox Release 2013a, The MathWorks, Inc.) ( Wang et al. 2013 ).

Figure 1.

Screen shot of the Matlab GUI during manual FP selection.

Figure 2.

Stained tongues with the FP (detected manually) demarquated with circles.

Automated selection of FP

Identification of stained tongue, and of circular markers

During the automated selection process, the white circular size markers were identified, followed by the identification of the blue stained area of the tongue. This was done by first converting the acquired image from RGB color-space to the more relevant HSV color-space. Because the white color remains invariant in the saturation value, the saturation channel was assigned a cut-off of 0.05 to generate a black and white image. To aid in the detection of circles, this black and white image was first shrunk to 20% of the original size, and further morphological closing was done using strel(‘disk’,10 ). Then the largest circles were detected using the function imfindcircles, with an upper and lower boundary on the radius as 30 and 90 pixels, objectpolarity as dark, and the sensitivity as 0.9 . The radius of the largest circle corresponds to 3mm in distance scale. These steps are depicted chronologically for one of the images, in Figure 3A–C . Next, the area of the stained region of the tongue was computed. This was done by first extracting the Hue channel from the HSV image ( Kuehni 2012 ), and generating a binary image for pixel values 0.55–0.59, which represents the color of the blue dye used to stain the tongue. These steps are shown for one of the images, in Figure 3D–F . To achieve the largest area that corresponds to the stained region, a further morphological closing was done using strel(‘disk’,10 ), followed by searching for the largest white region using bwlabel(X ). The identified stained tongue regions for each of the 9 images are shown in Figure 4 , with a red line marking the y axis corresponding to the lowest point in each image. The lowest point in the y axis of the stained area of the tongue was identified in order to find the distribution of FP along the vertical axis. Then the pixels were grouped into bins of 50 pixels in size, such that the first bin corresponds to the lowest point in the y axis.

Figure 3.

Figure showing the algorithmic steps to identify FP, by identifying the area of the size marker ( A – C ) and identifying the stained area of the tongue ( D – F ).

Figure 4.

Images showing the regions identified by the algorithm as stained regions, with the lowest point of each region marked with a line.

Calculation of the size, distance, and density characteristics

A range of properties were calculated, pertaining to the distribution and size of the FP. To normalize these values to physically measureable distances, the adhesive white circular markers, which were of uniform size and measured 6mm in diameter, were calculated in units of pixels. To calculate the area of the region representing the stained area of the tongue, the corresponding values of the saturation channel from the HSV colorspace were identified ( Cheng et al. 2001 ). This is due to the fact that in the Saturation channel, the intensity of the pixels representing the FP is lower than the surrounding area ( Liu et al. 2007 ). The resultant grey scale image was then passed through a morphological top-hat filtering followed by a bottom hat filtering. The structuring element used in these filters is defined as strel (‘ disk’,50 ) ( Serra 1982 ; Vandenboomgaard and Vanbalen 1992 ; Adams 1993 ). A hough transform based approach was then implemented, to identify circular features, corresponding to FP ( Yuen et al. 1990 ; Atherton and Kerbyson 1999 ), using imfindcircles with the radius of the FP limited between 20 and 50 pixels, a sensitivity of 0.90, and the “ObjectPolarity” chosen to be “Bright” ( Yuen et al. 1990 ; Atherton and Kerbyson 1999 ). The imfindcircles uses a phase-coding approach, using complex values in the accumulator array, and the radius information is encoded in the phase of the array entries. The function returns the centre locations corresponding to each individual FP, as well as the radius of the circle associated with the size of the FP ( Davies 2012 ). Figure 5 shows the automatically detected FP for the visible portion of the stained tongue regions for the 9 selected images.

Figure 5.

Stained tongues with the FP (detected algorithmically) demarquated with circles.

Comparison between the manually counted versus software-detected FP

To compare the software FP counts with those of a researcher, 2 circular regions left and right of tongue midline and close to the tip of the tongue were chosen, previously shown to correlate with the total number of FP on the tongue ( Shahbake et al. 2005 ). Researcher counts for each side were calculated, and an average taken. The same was done for the software-detected FP. This method simulated the between-researcher counts conducted in other studies, and the corresponding correlation between the 2 sets of averaged counts was very good-to-excellent, with an R² = 0.88 ( Figure 6 ).

Figure 6.

Correlation between manually detected ( x axis) and the algorithmically detected FP ( y axis) in the front of the tongue denoted by 2 manually drawn circles, using a sensitivity value of 0.92.

The images were used to quantify the total number of FP observed on the visible portion of the tongue, the average distance between FP, the size, and the density. The average Euclidean distance between FP was computed [for a total of x FP, there are x(x − 1)/2 possible unique distances], and the final distance in pixel units was averaged over all the unique distances. Here, the function pdist(X, ‘Euclidean’ ) was used to compute the Euclidean distance, where X is an x -by-2 matrix recording the x and y coordinates of x individual observations ( Press 1992 ; Deza and Deza 2009 ;). Then the Spearman correlation, averaged over x(x − 1)/2 unique points was calculated, which gives a measure of the degree of closeness between the FP. The function pdist (X,‘spearman’ ) was used to compute the Spearman correlation where the 2 sets are first ranked separately, and then the differences in rank are calculated at each position, i ( Fieller et al. 1957 ). The Spearman correlation distance between sequences X = ( X₁ , X₂ , etc.) and Y = ( Y₁ , Y₂ , etc.) was computed using the following formula

$$1-\frac{6{\displaystyle \sum _{i=1}^n{[rank(X_{\text{i}})-rank(Y_{\text{i}})]}^2}}{n(n^2-1)}$$

To compute the density in terms of number of pixels, the total stained area of the tongue was computed as above, and divided by the total number of pixels. To measure the area of the FP, a square region of 100 pixels around the FP was selected, corresponding to the values of the saturation channel from the HSV colorspace ( Ohta et al. 1980 ; Sural et al. 2002 ). The most prominent circle with an area that corresponds to the area of the FP was identified, using a hough transform based approach for identifying circular features, using imfindcircles , of radius between 20 and 50 pixels, a sensitivity of 0.95, and the “ObjectPolarity” chosen to be “Dark.” The prominence of the circle was measured as the highest value of metric variable returned by the imfindcircles function ( Sural et al. 2002 ). The greyscale saturation channel was then converted image to a black and white image using the function im2bw , with the graythresh function applying Otsu’s method, which chooses the threshold to minimize the intraclass variance of the black and white pixels ( Otsu 1979 ). The ratio of the black region returned by im2bw and the area of the circle, was used to define the roundness of the FP. Figure 7 shows the saturation channel for a sample of 8 FP, from the first image and the corresponding black and white images with the FP marked as green circles.

Figure 7.

Saturation channel images for 8 FP (taken from one image) of size 100×100 pixels, and their corresponding black and white (BW) images (FP marked with a a circle).

To find the distribution of the FP based on their radius, the radius of the most prominent circle was identified as described above. An estimation of the “roundnesss” of the FP was calculated via the distribution of the ratio of the black region (returned by the function im2bw, described earlier) to the area of the predicted circle. As an example of the characteristics that can be calculated using TongueSim, the distribution of the number of FP, the size, and roundness across the tongue are shown in Figure 8 (for the researcher-identified FP). The results of the number, size, and roundness are shown in Table 1 .

Figure 8.

Distribution of FP (researcher-identified) for 9 tongue images.

Table 1.

Parameters that are measurable by the software: distribution, size, and roundness of FP

Image number	Total FP per visible tongue	Average distance between FP centres (mm) a	Average size (mm 2 )	Average roundness of the FP
1	87	8.9	1.6	1.44
2	95	14.8	2.2	1.62
3	48	9.5	2.1	1.74
4	158	16.1	2.1	1.91
5	188	14.0	2.0	1.91
6	111	11.3	1.8	1.80
7	122	12.5	0.8	2.26
8	100	14.6	1.2	1.70
9	142	16.5	2.4	1.28

^a The size marker on each tongue was used to calculate the number of pixels per cm, and an individual conversion factor for each picture was calculated to convert pixel coverage to length estimates in mm and area estimates in mm ² .

Discussion

Here, we describe the development of software to automate the process of counting FP, and show good-to-excellent correlation with researcher counts, when used to count small regions on the tongue. Other inter-researcher correlations reported previously, for FP densities in these regions, range from 0.6 to 0.78 ( Feeney and Hayes 2014b ), demonstrating that this software works very well under the conditions used in the present analysis to identify FP. In addition to the reduction in time and error in taste research, other advantages include the measurement of other properties, such as degree of roundness of the papillae. There has been debate in the literature about the nature of the link between FP density and taste intensity perception, with some reporting that taste intensity is related to papillae density ( Delwiche et al. 2001 ; Doty et al. 2001 ) and others observing no association ( Garneau and Derr 2013 ; Feeney and Hayes 2014a ; Feeney et al. 2014 ; Garneau et al. 2014 ). A small clinical study of 10 people by Negoro et al. (2004) suggests that irregularly shaped papillae may be associated with reduced taste function. Melis et al. (2013) classified papillae distortion through the measurement of papillae diameters in 4 different rotations, and calculation of the standard deviation (SD). In that study, a papillae was considered distorted when the SD was 0.088 or greater, and the percentage of distorted papillae appeared to have a small, but significant, association with taste intensity of PROP in a cohort of 63 participants ( Melis et al. 2013 ). Given those findings, it is possible that the inclusion of irregularly shaped papillae with potentially reduced function could have resulted in the lack of association between FP and taste intensity that has been observed in the more recent, larger studies. However, there is currently a paucity of literature on this subject in relation to taste compounds other than PROP, and the calculation of distortion is time-consuming. The use of automated software with the capacity to estimate the degree of roundness in an objective manner may help to answer the remaining questions in this area. With taste function implicated in many conditions, including obesity ( Overberg et al. 2012 ), Parkinson’s disease ( Doty et al. 2014 ), and cognitive decline ( Steinbach et al. 2010 ), among others, we envisage that such software will have a wide variety of novel research applications.

The algorithms used here appear generally robust in the identification and measurement of FP when small, evenly stained regions are examined, as depicted in the images. However, it should be noted that are a number of situations where the algorithm may be less effective. For example, as shown in Figure 9A the markers used are not white in color. This results in impaired detection of the circular markers, and as a result the pixel per millimetre ratio is not well-predicted. Other situations whereby the algorithm may deliver low degrees of accuracy are in the event of uneven staining of the tongue. The algorithm makes the assumption that staining is uniform over the tongue surface. Examples are shown in Figure 9 where this is not the case, and as a result the algorithmic estimated total tongue area may be incorrect, resulting in an inaccurate estimate of density. Partial staining may also result in subsampling during the automatic detection of FPs (in other words, the software may return a lower number of FP than were actually there), and this can be seen in Figures 2 and 5, where the researcher-identified FP on the visible surface were not all identified by the software. For this reason, the method works best when comparing small regions rather than all FP from the visible portion of the tongue. In the event of nonuniform staining, as shown in Figure 9C the hough transform algorithm may over-estimate the number of FP. Therefore, caution is recommended in the interpretation of data resulting from nonuniformly stained images, or from images that estimate regions other than the small areas near the tip, as shown in the images. Nonetheless, the software presented represents a novel application for image analysis that may have benefits in future taste and FP research.

Figure 9.

Examples of various conditions where the algorithm may not fully identify FP: ( A ) circular marker not white in color, ( B ) staining is partial, and ( C ) staining is nonuniform.

Conclusions

Here, we describe novel software to automate the process of counting FP in a research setting. The correlation between researcher counts and the algorithmically detected FP was good-to-excellent, validating its use as alternative method to measure FP density via small regions near the tongue tip (caution is advised when estimating counts from regions other than these). Further, parameters such as FP roundness estimates and coverage can be relatively easily calculated, potentially extending the current repertoire of variables that can be readily assessed within oral anatomy research. Future research should address whether the degree of FP distortion is related to taste intensity, in a large data set.

Funding

No specific funding was obtained for this project. S.S. is supported by SimSci (Funded under the programme for Research in Third-level Institutions and cofunded under the European Regional Development Fund). E.L.F. is supported by Food For Health Ireland, FHI, through Enterprise Ireland (EI Grant no: TC-2013-001). The GIANT-CS study, from which the images were originally obtained, was supported by funds from the Pennsylvania State University, United States Department of Agriculture Hatch Project PEN04332 funds, and a National Institutes of Health grant from the National Institute National of Deafness and Communication Disorders [DC010904]. The authors would also like to thank all the research scientists involved in the data collection in the GIANT-CS study.