Developing a Robust Estimator for Remote Optical Erythema Detection

Maksym Ptakh

1,2

and Gennadi Saiko

Swift Medical Inc., 1 Richmond St. W, Toronto, Canada

University of Waterloo, Waterloo, Canada

Keywords: Erythema, Inflammation, Turbid Tissues, Optical Biopsy.

Abstract: Introduction: Erythema is redness of the skin or mucous membranes, which is symptomatic for any skin injury,

infection, or inflammation. In some cases, it can be indicative of certain medical conditions (e.g.,

nonblanchable erythema in Stage I pressure injuries), and its detection can facilitate intervention at an earlier

timepoint. The most common and effective means of erythema detection is a visual inspection of the skin.

However, in many cases (especially for people with darkly pigmented skin), erythema can be masked by

melanin. Moreover, it would be useful to have an automated delineation and measurement of erythema using

consumer-grade devices, e.g., smartphones. It would facilitate automated symptom detection and measuring

healing progress in various settings, including the patient's home. Aims: This study aims to evaluate and

compare several algorithms that can be used for automated erythema detection using a smartphone's camera

in clinical settings. Methods: We have compared three potential estimators, which can be derived from an

RGB image: a) log(R/G), b) R-G, and c) a* channel in CIELAB color space. Here, R and G are red and green

channels of an RGB image, respectively. Imaged skin was divided into two classes: erythema and non-

erythema. The "erythema" class was seeded with pixels with E>mean(E)+z*st.dev(E), where E is the value

of the estimator for a particular pixel, z is a model parameter (z-score). The erythema cluster was then grown

by gradually adding nearby regions with an estimator E closer to the estimator’s mean of erythema cluster

than the mean of the estimator for the normal skin area (K-Mean (K=2)). The segmentation algorithm was

tested on a subset of labeled images from the Swift Medical proprietary wound imaging database. To evaluate

algorithm performance, the results of segmentation were compared with ground truth, manually labeled

images. To quantify results, sensitivity, specificity, and ROC curves were used. Results: We have found that

all investigated estimators could provide reasonable sensitivity (>0.8) and specificity (>0.78). However, a*

based estimator offers slightly better performance (0.86/0.84). Discussion: The preliminary data shows that

smartphone cameras can delineate erythema with reasonable sensitivity and specificity. Further studies are

required to correlate the accuracy with the skin type (melanin concentration in the skin).

1 INTRODUCTION

Erythema is redness of the skin or mucous

membranes caused by hyperemia in capillaries. It is

symptomatic of any skin injury, infection, or

inflammation. In some cases, it can be indicative of

certain medical conditions (e.g., nonblanchable

erythema in Stage I pressure injuries), and its

detection can facilitate intervention at an earlier

timepoint. For example, detecting a Stage I ulcer will

allow timely intervention to prevent the ulcer's

progression.

https://orcid.org/0000-0002-5697-7609

The most common and effective means of

erythema detection is a visual inspection of the skin.

However, for people with darkly pigmented skin,

erythema can be masked by melanin. One specific

benefit of a robust erythema detection algorithm is the

development of an instrument for use by health care

professionals to detect erythema. This can be useful

in monitoring reactive hyperemia or detecting Stage I

pressure ulcers in intensely pigmented subjects.

Several techniques have been proposed to

increase the sensitivity and specificity of erythema

detection. Tissue Reflectance Spectroscopy (TRS) is

a non-invasive method of quantifying skin color. In

Ptakh, M. and Saiko, G.

Developing a Robust Estimator for Remote Optical Erythema Detection.

DOI: 10.5220/0010192901150119

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 2: BIOIMAGING, pages 115-119

ISBN: 978-989-758-490-9

115

particular, TRS has been used to characterize the

presence of erythema due to reactive hyperemia or

Stage I pressure ulcers (Hagisawa, 1994). While TRS

is a data-collection technique, the absorption data

have to be processed by an algorithm to detect and

quantify the erythema. In Riordan et al. (Riordan,

2001), five different algorithms have been compared.

The authors found that most algorithms demonstrated

adequate validity across all subjects. However,

spectroscopic techniques have certain limitations.

Firstly, they are a single point measurement, which

precludes them from providing additional clinical

parameters, e.g., redness surface size, which can be

used by dermatologists, alergologists, and other

clinical specialists. Secondly, it may require contact

with the skin, which is undesirable in many cases.

Finally, they are labor- and time-consuming and

require specialized equipment, which cannot be

universally available.

With the proliferation of smartphones and

improvements in their cameras, they have become

standard tools for healthcare professionals to measure

and document wounds and skin conditions. These

measurements are remote and non-invasive. More

importantly, they can be performed in any setting,

including the patient's home. Thus, the ability to

detect erythema using a smartphone can have a

significant clinical value.

This study aims to evaluate and compare several

estimators that can be used for automated erythema

detection using a smartphone's camera.

Skin detection and tissue type analysis are fairly

active research areas. Skin detection is important for

many applications (e.g., automated screening for

adult content detection). Tissue type analysis and

classification are important for wound care

applications.

These areas use multiple approaches, which

typically fall into a) traditional image processing

methods (e.g., Mukherjee et al. (Mukherjee, 2014)) or

b) Machine Learning (ML) algorithms, and

particularly deep neural networks (DNN) (e.g., Wang

at al. (Wang, 2015)). In some cases (see, for example,

Veredas et al. (Veredas, 2010) or Li et al. (Li, 2018)),

hybrid methods are used.

Skin detection and segmentation are well

performed using conversion into YCbCr color space

(see Brancati et al. (Brancati, 2017)). In YCbCr

space, skin colors for healthy skin are clustered in a

compact area, which can be approximated by an oval

(Hsu, 2002)).

Machine learning methods require labeled

images. While Swift Medical has its own database of

labeled wound images, in our first proof of concept

study, we did not use any ML approaches. The reason

for this is the following. While wound tissue types

(namely epithelial, granulation tissue, slough, and

eschar) can be considered "absolute," i.e., their colors

are independent of the color (tone) of the surrounding

skin, erythema colors are "relative" with respect to the

surrounding skin. Thus, wound tissue types are ideal

candidates for the ML, and particularly for DNN-

based algorithms. However, the "relativeness" of

erythema colors makes it possible to apply traditional

image segmentation techniques. Moreover,

traditional methods can be useful to derive and

quantify underlying physiological information.

While several attempts were made to develop and

analyze such classifiers before (e.g., Roullot et al.

(Roullot, 2005)), these studies were conducted in a

well-controlled lab environment on healthy

volunteers. While it is useful as a proof of concept and

benchmarking, it is not clear how these classifiers will

perform in real-life scenarios on patients with

wounds, dressings, etc. This article aims to evaluate

the performance of classifiers in a realistic setting on

wound care patients.

The article is structured as follows:

First, we discuss several potential estimators,

which can be derived from simple physiological

considerations.

Then, we discuss the cluster segmentation

algorithm to segment the erythema cluster.

Finally, we evaluate the estimators' performances.

2 METHODS

2.1 Estimators

We can try to select candidates for an erythema

estimator based on simple physiological

considerations. It is known that erythema is

characterized by an elevated blood supply. Thus, one

can expect that erythema will be accompanied by

reduced reflectance in the green range of the spectrum

(oxyhemoglobin absorption peaks) and

approximately the same tissue reflectance in the red

range of the spectrum (oxyhemoglobin absorption is

small).

Based on these considerations, we can consider

several potential candidates for estimators.

Diffey et al. (Diffey 1991) proposed 𝐸





log R





 . Here R

635

and R

525

are the

reflectances of the skin at 635nm and 525nm,

respectively. Based on this idea, we can start from the

BIOIMAGING 2021 - 8th International Conference on Bioimaging

116

following estimator based on red and green channel

pixel values

𝐸



log

𝑆



𝑆





(1)

Tronnier erythema index (Tronnier, 1969) is

based on the difference between red and green

reflectance at a control site and an erythematic area.

Melanin compensation is achieved by comparing two

sites. Based on these considerations, we can introduce

another estimator:

𝐸



𝑆



𝑆



(2)

Finally, we can take into account that in CIELAB

color space (Lab color space): L* is the lightness,

which changes from black (0) to white (100), a*

changes from green (−) to red (+), and b* changes

from blue (−) to yellow (+). Taking into account that

definition, we can transform the initial image from

RGB to Lab color space and use a* channel as an

estimator:

𝐸



𝑎

∗

(3)

2.2 Test Set

The estimators' performance was evaluated on the

wound images from Swift Medical (Swift Medical

Inc, Toronto, Canada) image repository. Swift’s

image repository consists of wound images taken by

a proprietary Swift Skin and Wound system using

iOS smartphone cameras. The image dimensions are

1077x808 and are in jpg format. 2000 images were

cleared of personally identifiable information (PHI).

Subsequently, images were labeled using a

browser-based image labeling platform (LabelBox)

by a team of trained labelers and reviewed.

Tissues were labeled using the following

categories: four types of wound tissue (epithelial,

granulation, slough, and necrotic),

maceration/erythema, normal tissue, a fiduciary

object, and other (e.g., gloves, cloth). For the

purposes of this pilot study, we manually went

through the dataset and selected a much smaller

subset, which contained the correct labeling of the

erythema. In particular, we selected 18 images that a)

contained erythema visually, and b) erythema was

correctly labeled, and 20 images that a) does not

contain erythema visually, and b) no erythema labels

on the image. An example of an unlabeled wound

image from the Swift Medical image repository is

depicted in Figure 1.

Figure 1: An example of an unlabeled wound image from

the Swift Medical image repository. The white/blue circle

at the center is a fiduciary object.

2.3 Cluster Segmentation

Each intact skin area was segmented into two

clusters: "normal" skin and erythema. Wound tissues,

fiduciary objects, and others (gloves, cloth) were

excluded from consideration (it was assumed that

other methods could identify these classes).

The segmentation algorithm consisted of the

following steps:

1. Find a "normal" skin cluster (manually or

automatically)

2. Calculate mean 𝐸





) and standard deviation

(𝜎𝐸



) for an estimator for all pixels within

the "normal" cluster

3. Seed an erythema cluster (R) using the

following algorithm

𝐸𝐸





𝑧𝜎𝐸





4. Grow the erythema cluster from seed points

using the algorithm similar to (Roullot, 2005):

a. Compute C, which is the region of pixels

adjacent to the current region R, obtained

with morphological dilatation: C = (R

⊕

) − R where

⊕

represents a

morphological dilatation with a 3x3

structuring element E

b. K-Mean (K=2) step. Compute C

, which

is the region of pixels that have an

estimator closer to the mean of R than the

mean of the normal skin area: C

= |E −

| > |E –



)| where E

is the mean

estimator over the area R

c. Update R : R = R

⋁

⋀

)

Developing a Robust Estimator for Remote Optical Erythema Detection

117

5. Repeat step 4 until R has no new pixels

6. To improve the accuracy of the results, the

noise was removed using open morphological

operations on R.

Here the first term 𝐸





) is the mean of the

estimator for the normal skin, 𝜎𝐸



 is the standard

deviation of the normal skin, z is a model parameter

(z-score).

The segmentation algorithm was applied to all

estimators under consideration.

2.4 Performance Evaluation

Labeled images were processed using the

segmentation algorithm for each estimator under

consideration and compared with the ground truth

(manually labeled images).

If the algorithm identified a pixel as erythema,

and it was labeled as erythema, then it was marked as

true positive (TP).

If the algorithm identified a pixel as erythema,

and it was not labeled as erythema, then it was marked

false positive (FP). If a pixel was not identified as

erythema; however, it was labeled as erythema, we

assign it to false negative (FN). Finally, if a pixel was

neither identified nor labeled as erythema, it was

marked as a true negative (TN).

Thus, for each image, we can calculate sensitivity

(true positive rate or TPR=TP/(TP+FN)) and

specificity (true negative rate or TNR=TN/(TN+FP)).

To find an optimal performance, we assessed

performance at different values of z-score and built

ROC (receiver operating characteristic) curves.

3 RESULTS

To compare estimators' performance, we calculated

sensitivity and specificity for several values of z (see

Table 1) and plotted ROC curves (see Figure2).

One can see that the a*-based estimator provides

the best prediction values. However, the performance

of diff (R-G) and log (log(R/G)) estimators follows it

closely.

4 DISCUSSION

Here we presented a pilot evaluation of potential

estimators, which can be derived from a regular RGB

image. While all estimators demonstrated reasonable

sensitivity and specificity, the a*- based estimator

outperformed the log(R/G) and R-G estimators. Thus,

transformation to another color space (namely,

CIELAB) provides some benefits. It also should be

noticed that results are relatively consistent in the

wide range of z-score (at least 1<z<3). It is a positive

sign, which indicates that it is probably not a spurious

finding.

The results are also in good agreement with

findings reported by other groups (Roullot, 2005).

A variety of factors can impact the accuracy of the

proposed approach. Firstly, various smartphones

have different color-correction mechanisms (auto

white balancing, AWB). Thus, disabling AWB can be

helpful to standardize colors. Secondly, the results

may be influenced by ambient illumination. Finally,

the comparison with ground truth can be problematic

for dark skin tones (e.g., V and VI). For example, it is

challenging to label erythema on dark skin. Other

means (for example, induced erythema) have to be

used instead of labeled images.

In future work, we plan to validate the algorithm

by studying the induced erythema on volunteers. In

particular, we plan to correlate algorithm

performance with skin tone. We also plan to compare

the performance of these estimators with CNN-based

classifiers.

Table 1: Performance of estimators at several z-scores.

Estimator log (R/G) R-G a*

Z-score Sensitivity Specificity Sensitivity Specificity Sensitivity Specificity

Z=1.0 0.811 0.767 0.836 0.78 0.862 0.836

Z=1.5 0.806 0.783 0.836 0.78 0.861 0.842

Z=2.0 0.806 0.783 0.836 0.781 0.86 0.842

Z=3.0 0.807 0.786 0.727 0.817 0.859 0.844

BIOIMAGING 2021 - 8th International Conference on Bioimaging

118

Figure 2: ROC curves for three estimators: R-G (blue

curve), log(R/G) (red curve), and a* (green curve).

5 CONCLUSIONS

We have analyzed the performance of several simple

estimators for erythema detection in realistic settings.

The preliminary data shows that smartphone cameras

can delineate erythema with reasonable sensitivity

and specificity. The approach can be implemented

using an inexpensive imaging setup (e.g.,

smartphone) and can be used in any setting.

ACKNOWLEDGEMENTS

The authors are thankful to Dhanesh Ramachandram

for help with the dataset.

REFERENCES

Hagisawa, S., Ferguson-Pell, M., Cardi, M., et al., 1994.

Assessment of skin blood content and oxygenation in

spinal cord injured subjects during reactive hyperemia.

J Rehabil Res Dev 31(1):1-14

Riordan, B., Sprigle, S., Linden, M., 2001. Testing the

validity of erythema detection algorithms. J Rehabil

Res Dev 38(1): 13-22.

Mukherjee, R., Manohar, D. D., Das, D. K., et al., 2014.

Automated tissue classification framework for

reproducible chronic wound assessment, BioMed

Research International, vol. 2014, Article ID 851582.

Wang, C., Yan, X., Smith X., et al., 2015. A unified

framework for automatic wound segmentation and

analysis with deep convolutional neural networks, in

Proceedings of the 2015 37th Annual International

Conference of the IEEE Engineering in Medicine and

Biology Society (EMBC '15), pp. 2415–2418.

Veredas, F., Mesa, H., Morente, L., 2010. Binary tissue

classification on wound images with neural networks

and bayesian classifiers, IEEE Transactions on Medical

Imaging, vol. 29, no. 2, pp. 410–427.

Li, F., Wang, C., Liu, X., et al., 2018. A Composite Model

of Wound Segmentation Based on Traditional Methods

and Deep Neural Networks, Computational

Intelligence and Neuroscience, vol. 2018, Article ID

4149103.

Brancati, N., De Pietro, G., Frucci, M., et al., 2017. Human

skin detection through correlation rules between the

YCb and YCr subspaces based on dynamic color

clustering, Computer Vision and Image Understanding,

vol. 155, pp. 33–42.

Hsu, R.-L., Abdel-Mottaleb, M., Jain, A. K., 2002. Face

detection in color images, IEEE Transactions on

Pattern Analysis and Machine Intelligence, vol. 24, no.

5, pp. 696–706.

Diffey, B.L., Farr, P.M., 1991, Quantitative aspects of

ultraviolet erythema. Clin Phys Physiol Meas

12(4):311-25

Tronnier, H., Evaluation and measurement of ultraviolet

erythema. In: F Urbach (editor). Biologic effects of

ultraviolet radiation. Oxford: Pergamon Press; 1969. p.

255-66.

Roullot, E., Autegarden, J.E., Devriendt, P., et al., 2005.

Segmentation of Erythema from Skin Photographs for

Assisted Diagnosis in Allergology. International

Conference on Advances in Pattern Recognition, p.754-

763.

Developing a Robust Estimator for Remote Optical Erythema Detection

119