Comparative analysis of two methods for wound bed area measurement

Abstract

Wound bed area measurements are considered to be an essential part of the wound assessment process. Wound care professionals should be aware of the reliability and validity of the techniques they use. The purpose of this study was to assess whether wound care professionals are able to make as accurate and reproducible a measurement of the wound bed area using two methods for area measurement. Five wound care professionals independently assessed 2285 digital wound images for the wound bed area. Each image was measured in random order, three times, and in four angles by providing the rotated versions of each image (0°, 90°, 180° and 270°). Two techniques were compared: free hand drawing and closed polygon (CP) graph algorithm. Comparison of the two techniques showed differences that are, in our opinion, not acceptable in clinical practice when these techniques are used interchangeably and/or the measurements are carried out by different observers. Variations observed between wounds and observers seem related to the difference in perception of the wound bed margin. Our results indicate that repetition of CP graph area measurement results in the lowest difference in repetitive measurements. Study limitations are related to an incomplete consensus on definitions of wound, wound bed, wound edge and wound border. The development of an ontology related to wound images could aid to reduce these ambiguities.

INTRODUCTION

Professionals dealing with wound patients make clinical decisions principally but not solely based on their visual perception 1, 2, 3, 4. The descriptive analysis of wounds, however, is still poorly standardised and rarely reproducible 5, 6, 7, 8, 9, 10, 11, 12. Measurement of time‐related changes based on digital images of wounds is optimised by new developments in calibration techniques 13, 14, 15. With a growing demand for randomised clinical trials in chronic wound care 16, 17, 18 and an increasing economical pressure on health budgets 19, 20, a key requirement for optimal data sharing is standardisation with agreements on the definitions of structures, processes and formats used. Although new camera systems deliver ever higher resolutions, a full understanding of cutaneous imaging techniques continue to deserve attention 21, 22. It seems generally accepted, albeit with an understanding for the legal underpinnings of its use, that imaging techniques can provide additional information and can assist in the management of skin problems 23, 24. They frequently offer the promise of automation, objectivity and reproducibility and have potentially higher sensitivity than expert human observers 25, 26. This article describes the degree of disagreement and/or agreement between two methods of wound area measurement when wound areas are measured by different observers.

This experiment is part of the research presented by the Woundontology Consortium, which is a semi‐open, international, virtual community of practice devoted to advancing the field of research in non invasive wound assessment by image analysis, ontology and semantic interpretation and knowledge extraction (www.woundontology.com). The interests of this consortium are related to the establishment of a community‐driven, semantic content analysis platform for digital wound imaging with special focus on wound bed surface area and colour measurements in clinical settings.

Digital imaging

Wound conditions can be monitored chronologically using digital imaging. There has been a growth in mathematical models designed to explain and extend the understanding of wound healing (27) and to analyse colours 28, 29, 30, patterns (31) and shapes of all sorts of objects. All approaches have the same goal in mind: disambiguation faced during interpretation of and communication on images.

Calibration

Digital images of human wounds with a reference chart [the Mac‐Beth ColorChecker Chart Mini (MBCCC) (GretagMac‐Beth AG, Regensdorf, Switzerland), as previously published by Vander Haeghen and Naeyaert (13), or the QP Card 201 (QPcard AB, Goteborg, Sweden)] were uploaded via a smart client tool (www.colibrate.com) to the image server. The illumination of the reference chart was homogeneous over the field of view. The colour and planimetric calibration algorithm used in this study is based on three one‐dimensional look‐up tables and polynomial modelling (Figure 1) (13). Recently, an automatic colour and planimetric calibration algorithm that ensures reproducible colour content and size of digital images was published (14). Evidence was provided there that images taken with commercially available digital cameras can be calibrated independently of any camera settings and illumination features.

Figure 1.

Colour and planimetric calibration algorithm.

Wound bed surface area

Previously, it has generally been recognised that there is no method of wound measurement which is accurate, repeatable, inexpensive and practical for everyday clinical use (32). Assessing the wound bed surface area change over time has been used, although it remains only a surrogate endpoint for healing (33). Each wound bed area measurement method has its own drawbacks. Consequently, using them in routine clinical practice is sometimes difficult or impractical. Simple and cost‐effective methods of wound assessment are required. Although segmentation algorithms have been designed for wound assessment, they are not commonly used on a day‐to‐day basis (34). The least sophisticated computerised method is manual segmentation, whereby an operator on a computer workstation freely draws a boundary on an image to indicate the location of the wound bed. The most widely used methods for measuring wounds are linear measurements of length and width (35), wound tracing and computer‐based or digital planimetry 36, 37. The perimeter is calculated as the length of the line that bounds the wound bed area.

Before a tool can be applied in clinical practice, a validation process that assesses the reliability and repeatability of its use by different health care professionals should be performed plus, in particular, in this setting, a specification of the region of interest (wound bed) should be defined. The purpose of this work is to evaluate the degree of disagreement and/or agreement of two basic screen‐based tools to measure the wound bed surface area from digital wound images in absence of a computer‐understandable definition of the wound bed. The first technique, the free hand (FH) drawing, can be used by simply holding down on the mouse button and dragging to draw the margin of the wound bed. The second technique, based on a closed polygon (CP) graph algorithm is a technique where the margin of the wound bed is drawn with multiple lines that eventually meet.

MATERIALS AND METHODS

Human subjects

The sample comprised patients (n = 16) with different chronic wounds (n = 20), recruited in an outpatient wound care facility (Ziekenhuis Oost‐Limburg, Belgium). The patients were in no way representative of any defined population. Only anonimised images were used. Conforming to the ethical guidelines of the 1975 Declaration of Helsinki, and with ethical approval, patients were given explanation of the study, and an information leaflet was provided. The patients were asked to sign a consent form if they agree to participate. The digital images (n = 19, one with two different wounds), considered as representational artefact of the corresponding wound, were preprocessed for analysis (colour and planimetric calibration), as previously published 13, 14, 38. The planimetric calibration (correlated to the fixed and known size of the reference chart) reduced the bias related to magnification and resolution of images under consideration.

Procedure

Following removal of the dressing, the wound care professional was asked to take a digital image of the wound. Consequently, the images were uploaded to the Woundontology Consortium image server platform on www.colibrate.com. Determining wound surface areas in this experiment involved a two‐stage process: identifying the wound bed margin using digital photography and calculating the surface area (3). Five health care professionals, with advanced and equal wound care experience, independently assessed 19 digital wound images (one of the images had two wounds) for the wound bed surface area three times and in four angles by providing the rotated versions of each image in random order (0°, 90°, 180° and 270°). Wound types varied, although mainly pressure ulcers and foot ulcers have been measured in the clinic where the study took place 39, 40, 41.

Statistics

Inter‐ and intra‐observer measurement variations were estimated through statistical modelling. The various sources of variation for the wound bed area and perimeter measurements were assessed with the analysis of variance. The model had five factors: wounds (n = 20), observers (n = 5), technique (n = 2), rotation (n = 4) and replication (n = 3). The model tested for differences in the average wound bed area and perimeter size among observers, the average wound area and perimeter size between the three replications, the two techniques and among the wounds. Repeatability within each wound measurement method was investigated by calculating a coefficient of variation (CV) for each wound measurement. Using Medcalc (version 10·0·1·0 for Windows), the Kruskal–Wallis test (H‐test) was used to determine whether any method was consistently more repeatable than another. A value of P < 0·05 was considered statistically significant. In order to compare the two strategies, the wound bed surface area data were plotted with line of equality (42). Consequently, the correlation was calculated between the two methods using a correlation coefficient and the Bland and Altman method. The null hypothesis was that the measurements by the two methods are not linearly related. A plot of the difference between the methods against their mean displays the degree of agreement between the two methods. The true value of the wound bed surface area was unknown, and the mean of the two measurements was considered to be the best estimate available. The distribution of the surface area and perimeter measurement was tested for normality using the D’Agostino–Pearson test computing a single P‐value for the combination of the coefficients of skewness and kurtosis (43).

RESULTS

Sample data

A total of 2285 wound images, representing different wound types, have been analysed. The principal characteristics of the image area (cm²) and perimeter (cm) evaluation are shown in Tables 1 and 2.

Table 1.

The principal characteristics of the image area (cm²) evaluation (kurtosis is a measure of whether the data are peaked or flat relative to a normal distribution)

Variable		Area
Sample size		2285
Lowest value		0·4170
Highest value		46·1822
Arithmetic mean		11·8333
95% CI for the mean		11·2892–12·3773
Median		5·4011
95% CI for the median		5·2162–5·5229
Variance		175·8694
SD		13·2616
Relative SD		1·1207 (112·07%)
Standard error of the mean		0·2774
Coefficient of skewness		1·4135 (P < 0·0001)
Coefficient of kurtosis		0·4592 (P = 0·0002)
D’Agostino–Pearson test for normal distribution		Reject normality (P < 0·0001)
Percentiles		95% CI
2·5	0·6412	0·6095–0·7119
5	1·4331	0·8285–1·5280
10	1·7181	1·6555–1·7959
25	4·1663	4·0767–4·2296
75	11·0350	10·3048–21·2066
90	39·8402	39·4046–40·0843
95	41·2140	40·8367–41·5177
97·5	42·0610	41·7914–42·3696

CI, confidence interval.

Table 2.

The principal characteristics of the image perimeter (cm) evaluation

Variable		Perimeter
Sample size		2285
Lowest value		3·0782
Highest value		42·3190
Arithmetic mean		13·3093
95% CI for the mean		13·0089–13·6096
Median		10·7831
95% CI for the median		10·5818–10·9692
Variance		53·6118
SD		7·3220
Relative SD		0·5501 (55·01%)
Standard error of the mean		0·1532
Coefficient of skewness		0·8493 (P < 0·0001)
Coefficient of kurtosis		−0·4588(P = 0·0002)
D’Agostino–Pearson test for normal distribution		Reject normality (P < 0·0001)
Percentiles		95% CI
2·5	3·8633	3·7204–4·0068
5	4·6015	4·3230–4·8793
10	5·3348	5·1885–5·6287
25	8·6522	8·5235–8·8032
75	19·2079	18·6458–20·0935
90	24·7566	24·3568–25·1546
95	28·3463	26·8934–28·6002
97·5	29·3094	29·0131–29·4735

CI, confidence interval.

Method comparison

The data, area and perimeter measurements using FH and CP and the line of equality (on which all points would lie if the two methods gave exactly the same reading every time) are represented in 2, 3. For the data in 2, 3, the correlation coefficient (r) between the two methods for area measurement was r = 0·99 (P < 0·001) and for the perimeter measurement: r = 0·99 (P < 0·001). However, a high correlation does not mean that the two measurements agree!

Figure 2.

Correlation of area measured with free hand and closed polygon with line of equality.

Figure 3.

Correlation of perimeter measured with free hand and closed polygon with line of equality.

The difference between the measurements obtained by the two different methods was found to be dispersed around the mean, with no clear trend towards over‐ or underestimation by either one of these measurement techniques, as graphically seen through the 95% agreement limits of Bland and Altman method (4, 5). It is unlikely that the two methods will agree exactly. The plot of difference against the mean allows us to investigate any possible relationship between measurement error and the true value. Because the true value was unknown, the mean of the two measurements is the best estimate available. Provided differences within the mean difference ± 1·96 SD of the differences would not be clinically important, the two measurement methods could be used interchangeably (limits of agreement). For the area and perimeter data of mean differences ± 1·96 SD were found, respectively, to be +1·1 and −1·1 cm² and +1·4 and −1·6 cm. These differences seem acceptable for clinical purposes, considering that the largest differences are positioned in the largest wounds.

Figure 4.

(A) Bland and Altman plot of area. (B) Bland and Altman plot of perimeter.

Figure 5.

(A) Bland and Altman plot of area percentage. (B) Bland and Altman plot of perimeter percentage.

However, if the differences as percentage of the averages were calculated, which is useful when an increase in variability of the differences is expected as the magnitude of the measurement increases, the largest percentage of differences was observed in the smaller wounds. At least one method depends strongly on the magnitude of measurements. For the area and perimeter data showed mean differences ± 1·96 SD of the differences, respectively: +10·5% and −9·6% for the area data and +10·5% and −11·4% for the perimeter data. These differences might not be acceptable for clinical practice when both techniques are used interchangeably. Both area and parameter measurements seem susceptible for these differences.

Inter‐observer variability

The inter‐observer CV for the area and perimeter measurements from all measured wounds expressed as percentage is, respectively, illustrated in Figure 6A, B.

Figure 6.

(A) Coefficient of variation for area measurement, expressed as percentage. (B) Coefficient of variation for perimeter measurement, expressed as percentage.

The mean, minimum and maximum inter‐observer CV for the area and perimeter measurements for all measured wounds expressed as a percentage are, respectively, 5·92%, 1·19% and 64·07% for the area data and 5·07%, 1·14% and 34·17% for the perimeter data. The mean, minimum and maximum inter‐observer CV for the area and perimeter measurements between the FH and CP techniques, expressed as a percentage are, respectively, FH: 5·85%, 1·01% and 63·16% for the area data and 5·25%, 0·72% and 36·67% for the perimeter data and CP: 5·81%, 1·16% and 67·66% for the area data and 4·53%, 0·63% and 33·03% for the perimeter data (Figure 7). The large variations during observation are illustrated in wound 40 and 70 and seem related to the difference in perception of the wound bed margin. Differences between real differences compared with measurements expressed as percentage might reduce the bias of wound area size. Although the influence of outliers is included in the calculation, it only shows the variations observed with different observers and different measurement techniques. Exclusion of the outliers would perhaps better illustrate the intrinsic properties of the respective methods.

Figure 7.

The mean, minimum and maximum inter‐observer coefficient of variation for area and perimeter measurements.

Intra‐observer variability

The intra‐observer variability was tested for each repetition and between the two techniques. Measurement results were normalised and compared. Kruskal–Wallis test (H‐test) was used to test the variability between the second and third area measurement. If the null hypothesis, being the hypothesis that the samples of the second and the third measurement originate from the same population (as the first measurement), is rejected (P < 0·05), then the conclusion is that there is a statistically significant difference between at least two of the subgroups. As shown in Figure 8, with P‐values of 0·0171 and 0·0148, a significant difference is found between the first and second measurement for both techniques together and FH, respectively. For the CP technique, a P‐value of 0·5333 suggests no statistically significant difference between the second and third area measurements. The perimeter measurement provided P‐values of 0·3975, 0·7114 and 0·3712, respectively, for both techniques, FH and CP.

Figure 8.

Difference between the first and second measurements for both techniques.

DISCUSSION

Comparison of both the techniques showed differences which are, in our opinion, not acceptable for clinical practice in the context that both techniques are used interchangeably and wound bed areas are measured by different wound care professionals.

This experiment also illustrates that a high correlation does not automatically mean that the measurements agree! Therefore, demonstrating difference between measurements by the technique described by Bland and Altman (42) might be more illustrative (Figure 4A, B).

Our results indicate that in this experiment for intra‐observer variability, the CP measurement results in the lowest difference in repetitive measurements.

The analysis was carried out to critically appraise the two‐dimensional methods of wound surface area measurement without taking in account volume changes over time and the curvature of the body. To date, only five participants have been involved which could be responsible for a low power of statistical significance. Overall usability is important for acceptance of a new technology, as potential users have to be convinced that it offers real benefits in terms of time saved and practicality and that it could result in higher quality patient care and cost savings. This usability was not tested in this article.

During each measurement, no guarantee was incorporated that identical values for wound bed area represented the same region.

More study limitations are related with an incomplete consensus on definitions of wound, wound bed, wound edge and wound border. The development of an ontology [a representational artefact whose representational units (which may be drawn from a natural or from some formalised language) are intended to represent (i) universals in reality and (ii) those relations between these universals which obtain universally (for all instances)] as a declarative model of the domain of wound care (imaging) that defines and represents the concepts existing in that domain, their attributes and the relationships between them could aid to reduce ambiguities. In an attempt to reduce these ambiguities, the discussion remains open on which parameters best describe the time‐based changes in wounds: wound border, wound bed colour and/or wound bed texture.

In pigmented skin lesions, border detection is often the first step in the automated analysis of dermoscopy images. Although numerous methods (e.g. the normalised probabilistic rand index) have been developed for the detection of lesion borders, very few studies were comprehensive in the evaluation of their results (44). Others found a good concordance between clinicians and computer evaluation during evaluation of the ABCD rule for dermoscopy (45).

The large variations observed between wounds and between observers seem additionally related to the difference in perception of the wound bed margin. The lack of a uniform, quantitative definition of the wound bed might explain these differences. In the challenge of quantification and semi‐automated determination of the wound bed area and border, further research could be based on previous definitions for different types of boundaries, including both the bona fide boundaries which we find in the physical world and the fiat (or human demarcation‐induced) boundaries (46).

ACKNOWLEDGEMENTS

We thank all the wound care professionals and participants who contributed to this study, in particular, the support and assistance of the Provinciale Hogeschool Hasselt and the University of Hasselt are greatly appreciated. All authors declare no conflict of interest, real or perceived, financial or non financial.