These articles have been accepted for publication in the British Journal of Dermatology and

are currently being edited and typeset. Readers should note that articles published below have
been fully refereed, but have not been through the copy-editing and proof correction process.
Wiley-Blackwell and the British Association of Dermatologists cannot be held responsible
for errors or consequences arising from the use of information contained in these articles; nor
do the views and opinions expressed necessarily reflect those of Wiley-Blackwell or the
British Association of Dermatologists
This article is protected by copyright. All rights reserved.

Received Date : 22-Sep-2014

Revised Date : 28-Nov-2014

Accepted Date : 20-Jan-2015

Article type : Clinical and laboratory investigations

Pilot Study of an Automated Method to determine Melasma Area and Severity Index

E.Y. Tay1, E.Y. Gan1, V.W.D. Tan1, Z. Lin2, Y. Liang2, F. Lin3, S. Wee2, T.G. Thng1
1
National Skin Centre, Singapore
2
School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore
3
School of Computer Engineering, Nanyang Technological University, Singapore

Corresponding Author:

A/Prof Steven Thng

National Skin Centre

1 Mandalay Road

Singapore 308205

Telephone: (65) 6253 4455

Fax: (65) 6253 3225

Email: steventhng@nsc.gov.sg

Running head: Automated Method to Determine Melasma Area and Severity Index (53 characters)

Conflicts of Interest: The authors have no conflicts of interest to declare.

This article is protected by copyright. All rights reserved.

Sources of Funding: A*STAR-NHG-NTU Skin Research Grant 2014 (SRG\14011)

What’s already known about this topic?

- The MASI and mMASI are the only 2 validated scores for assessing melasma severity. They
are essential to the evaluation of treatment response in trials and in the clinical setting.
- However, MASI and mMASI scoring is subject to much inter -observer variability and training
is necessary to ensure consistent scoring.
-
What does this study add?
- A novel image analysis software has been developed to derive automated MASI scores.
- Automated mMASI scores correlate well with clinician scored mMASI scores.

Summary

Background
Objective outcome measures for melasma severity are essential for the evaluation of severity as well
as results of treatment. The modified MASI score (mMASI) is a validated tool for assessing melasma
severity but is often subject to inter-observer variability.

Objectives
To develop and validate a novel image analysis software designed to automatically calculate the area
and degree of hyperpigmentation in melasma from computer image analysis of whole-face digital
photographs, thereby deriving an automated mMASI score (aMASI).

Methods
The algorithm was developed in collaboration between dermatologists and image analysis experts.
Firstly, using adaptive threshold method; the algorithm identifies, segments and calculates the areas
involved. It then calculates the darkness. Finally, the derived area and darkness are then used to
calculate mMASI. The scores derived from the algorithm are validated prospectively. Twenty-nine
melasma patients on depigmenting agents were recruited for validation. Three dermatologists
scored mMASI at baseline and post-treatment using standardised photographs. These scores were
compared to aMASI scores derived from computer analysis.

Results
aMASI scores correlated well with clinical mMASI pre-treatment (r=0.735, p <0.001) and post-
treatment (r=0.608, p<0.001). aMASI was reliable in detecting changes with treatment. These
changes in aMASI scores correlated well with changes in clinician-assessed mMASI (r=0.622,
p<0.001).

Conclusions
This study proposes a novel approach in melasma scoring using digital image analysis. It holds
promise as a tool that would enable clinicians worldwide to standardise melasma severity scoring
and outcome measures in an easy and reproducible manner, enabling different treatment options to
be compared accurately.
This article is protected by copyright. All rights reserved.

Pilot Study of an Automated Method to determine Melasma Area and Severity Index

Introduction

Melasma is an acquired disorder of hyperpigmentation which commonly affects women of

Fitzpatrick skin type IV to VI1. It is a common presenting complaint in dermatology clinics and affects
more than 5 million people in the United States2,3. In Southeast Asia, the prevalence reaches 40% in
women and 20% in men4. As a chronic and often relapsing condition, it has significant negative
impact on quality of life, even more so than in dermatoses such as acne and rosacea5. Melasma is
therapeutically challenging and new treatment options are constantly being explored. However, a
common obstacle faced is the lack of good and sound outcome measures, which is a critical but
difficult-to-standardise aspect in clinical trials.

The Melasma Area and Severity Index (MASI) and modified MASI score (mMASI) are the most
frequently used instruments to measure disease severity and treatment response in clinical trials.
The MASI (Fig. 1) is calculated based on the darkness, homogeneity and extent of pigmentation 6. The
MASI has been validated and performed well, when compared with the categorical Melasma
Severity Scale and mexameter readings. The validated modified MASI score (mMASI), which removes
homogeneity as a scoring component, has been shown to be as reliable as MASI scoring7.

While reliable, both MASI and mMASI have their distinct disadvantages. Firstly, consistency is
difficult to achieve. In practice, evaluators assign MASI and mMASI scores with substantial variability.
As MASI is based on subjective visual assessment of area of involvement and darkness, untrained
evaluators may over- or underestimate the surface area of involvement as well as how dark it is. This
results in inter-assessor variability. Assessment of the chin, which makes up 10% of the mMASI
score, is particularly difficult7. Secondly, the area and darkness component of the score are grouped
in bands, this is not ideal as significant changes within each band will not be reflected, resulting in a
falsely negative unchanged score. Conversely, small changes across bands result in a significant
change in scores. Finally, MASI is rarely used outside clinical trials due to the complexity of
calculations, its time-consuming nature and its inherent lack of intuitive meaning compared to
simple measures such as surface area of involvement.

To circumvent this, we developed a novel computer image analysis software to derive the extent
and degree of hyperpigmentation in melasma. These scores are fed into the same formula used for
mMASI scores cumulating in an automated MASI score (aMASI). We then validated the aMASI in a
pilot study involving 29 subjects, to evaluate the reliability of this software in the scoring of melasma
in response to treatment, as compared to the conventional method of clinician scored mMASI.

Methods

Developing the Algorithm.

The algorithm was developed in collaboration between dermatologists from National Skin Centre,
Singapore and computer science engineers from Nanyang Technological University, Singapore. To
derive the score, 3 photographs are needed, one full frontal picture for calculating the forehead and
the chin component, one each from the left and right side of face to calculate the malar regions. The
algorithm uses 3 main steps to derive the aMASI scores.
This article is protected by copyright. All rights reserved.

The first step is identifying the areas involved. To achieve this, an area of normal skin is manually
selected. This area would be used as a reference point. The colour difference between each point
and the reference point would be calculated. If this value is larger than a certain threshold, the point
will automatically be classified into the non-skin set. Using this method, the pixels of the eyebrows,
eyes, nostrils and corners of mouth will automatically fall into the non-skin set. The non-skin pixels
are then automatically programmed to be black and thus excluded from the analysis (Fig.2). The
identified areas are then segmented. We use the position of the eyes, nostrils and mouth as
landmarks to determine the boundaries between the forehead, malar and chin regions. From the
frontal view, any area above the horizontal line that passes through the eyes would be counted as
part of the forehead. Using the frontal view, the area beneath the curve that passes the centre of
the nostrils and lateral borders of the mouth would be demarcated as the chin area (Fig. 3a). From
the lateral view; the malar region is defined superiorly by a line drawn from the eyelid to the frontal
hairline, inferiorly by a curve from the corner of the mouth to the nostril, and laterally by the jawline
(Fig. 3b). Thresholding is then performed. This is done by comparing the value of one pixel against
the reference value of normal skin. Any area that is darker than the reference value would be
considered as an area affected by melasma (Fig.4). The area involved is calculated as a percentage of
the total surface area of the segment of the patient’s face that is being calculated

Subsequently, the darkness of the segmented area is calculated. This is done by determining the Lab
values of normal and affected skin on the patient’s face using the Lab color space8. The difference
between the Lab values of normal and affected skin will be calculated for each pixel. To get the
overall darkness in a specified area, we find the average darkness of the pixels in areas affected by
melasma.

Finally, the area involved and darkness component will be fed into the formula for mMASI (mMASI =
0.3 A(f)D(f) + 0.3 A(rm)D(rm) + 0.3 A(lm)D(lm) + 0.1 A(c)D(c), where f=forehead, rm= right malar, lm=
left malar, c = chin, A= percentage area involved and D= darkness) to generate an automated MASI
(aMASI) score (see Fig. 1).

Validating the Algorithm

We conducted a prospective study of 29 subjects with melasma, who received standard

depigmenting agents from October 2013 to March 2014. The study was approved by the
Institutional Review Board (Protocol 2014\0032).

Each subject underwent standardised photography at baseline and post-treatment. Written

informed consent was obtained prior to photography. A frontal view and 2 lateral views with 45°
rotation to the left and right were obtained at the same distance and using a constant light source
with the Visia™ Complexion Analysis system (Canfield Scientific Inc., Fairfield, NJ, USA). The eyes,
hair and lips of the subjects in these photographs were digitally erased so as to maintain patient
confidentiality.

In calculating the mMASI, the face was divided into 4 separate and mutually exclusive regions:
forehead (f), right malar (rm), left malar (lm) and chin (c); corresponding to 30%, 30%, 30% and 10%
of the total face respectively. The area of involvement (A) in each of these 4 areas was given a
numeric value of 0 to 6 (0= no involvement; 1 = <10%, 2= 10-29%; 3= 30-49%; 4= 50-69%; 5= 70-89%,
6= 90-100%). Darkness (D) was rated on a scale from 0 to 4 (0= absent; 1= slight, 2= mild; 3=marked;
This article is protected by copyright. All rights reserved.

4=maximum). The mMASI was given by the sum of the severity ratings for darkness multiplied by the
weighted value of the area involved for each of the 4 areas (see Fig.1). This resulted in a minimum
score of 0 and maximum score of 247.

The clinical mMASI score was derived from baseline and post-treatment photographs by a senior
consultant, consultant and resident in dermatology. The clinicians scored independently of each
other.

The same set of photographs that were used by the clinicians to score the melasma were fed into
the developed algorithm to derive the aMASI score. An area on the digital photograph was selected
by the user as the normal area before running the algorithm.

In order to detect a correlation of 0.6 between mMASI and aMASI at the 5% significance level (2-
sided) with a power of 90%, a sample size of 29 patients was calculated to be sufficient. Statistical
analysis was performed using SPSS software, version 11.0 (SPSS Software, Chicago, Il, USA).
Validation of aMASI against clinical mMASI was by Spearman’s correlation coefficient. Statistical
significance was set at p<0.05.

Results

For validation, 29 patients with melasma were recruited. Twenty-six were females (%) and 3 were
males (%), with a mean age of 48.6 years. All patients had Fitzpatrick Type IV to VI skin type.

The clinical mMASI and aMASI scores are shown in Table 1. The mean clinical mMASI at baseline was
3.64 ± 2.45, while the mean clinical mMASI post-treatment was 3.16 ± 2.26. The mean change of
clinical mMASI was -0.476 ± 0.87. In comparison, the aMASI at baseline was 3.54 ± 1.31. Post-
treatment, the mean aMASI was 3.33 ± 1.42. This corresponded to a mean change in aMASI of -0.21
± 0.67.

For the baseline images, Spearman’s correlation coefficient (r)= 0.735, showed that automated
aMASI correlated well with clinical mMASI (p<0.001) (Fig. 5). This was similar for imaging post-
treatment (r= 0.608, p<0.001) (Fig. 6). The change in aMASI scores post treatment were also
compared with the clinical change in mMASI scores and these also correlated well (r= 0.622,
p<0.001) (Fig. 7).

Discussion

The MASI and mMASI are considered the most valid and widely used instruments in the assessment
of melasma. The use of MASI in clinical trials has allowed assessment of melasma severity states and
is undeniably a useful measure of response to treatment. However, similar to the Psoriasis Area and
Severity Index (PASI) from which the scoring of MASI was derived, it is subject to much inter-rater
variability9. Studies have also shown that the PASI does not discriminate well among patients with
reduced extent of involvement, which can similarly be extrapolated to the MASI scoring10.

This newly developed algorithm takes the variability out of the MASI score by leveraging on digital
analysis of clinical photographs to derive the individual components of darkness and extent of
involvement. As this is done semi-automatically, minimal training is needed from the assessors. The
This article is protected by copyright. All rights reserved.

only manual step involves selection of skin unaffected by melasma prior to the running of the
algorithm.

In addition to assessor variability, the clinical MASI score groups area involvement as well as
darkness into bands. This is disadvantageous, as changes within the band (for example, if the area
treated has reduced from 60% to 50%) are not reflected at all in the final score even if the patient
has clinically improved. Similarly, small changes across bands (i.e reduction of area involved from
70% to 65%) results in a big change in MASI score due to the artificiality in banding. The developed
algorithm overcomes this limitation as the scores are not banded. All changes, whether large or
small, will be reflected accordingly in the scores.

Finally, the algorithm is simple and easy to use. In the future, we also plan for it to be downloaded
via the internet so that a standardised scoring for melasma that can be made available to
researchers worldwide. There is no need for training and the algorithm will generate consistent
aMASI scores that can enable clinicians around the globe to standardise melasma severity scoring
and outcome measures in an easy and reproducible manner.

A novel aspect of the algorithm is the combined use of global and adaptive threshold segmentation
technique to select the area affected by melasma11-13. To our knowledge, such a combination has not
been used in the imaging of patients with melasma. This combination enables us to overcome the
problem of determining areas affected by melasma, as the boundary between melasma and
unaffected skin is often not clearly defined and blurred by diffusion.

While useful, this algorithm has the following weaknesses. Firstly, most of the patients in our cohort
had mild melasma, which is a reflection of the severity of melasma of the majority of patients seen
in our centre. This algorithm would need to be prospectively validated in patients with moderate to
severe melasma. Secondly, while the algorithm is able to identify subtle differences between
pigmented skin and normal skin, these detected variations may not be due to melasma. As of now,
this algorithm is unable to distinguish between melasma and other forms of hyperpigmentation such
as Hori’s nevus and solar lentigenes. It is not unusual for patients with melasma to have concomitant
hyperpigmentary disorders which can be more easily distinguished by clinicians than the automated
system. In addition, male patients with moustaches may need to have their moustaches be shaved
before the photographs are taken or the scores generated by the computer may be affected. Thus,
at present, the algorithm is best suited to patients who neither have any concomitant pigmentary
disorders nor an excess of facial hair.

Moreover, digital analysis of the chin for aMASI scoring is difficult as the chin area has a tendency to
be obscured in photographs. In addition, we note that proper selection of normal skin is also
essential to ensuring the consistency of aMASI scores. If the illumination is not uniform and the area
of normal skin selected is over- or under-illuminated, the results may be skewed. Finally, all the
patients in this study were of Fitzpatrick skin types III to VI and further validation in Fitzpatrick skin
types I and II would be needed.

The current validation of the algorithm is a small pilot study, and we plan to validate this aMASI
algorithm in a larger prospective cohort where clinical mMASI will be scored based on actual clinical
examination rather than with photographs. This algorithm would also need to be validated
prospectively by multiple centres and in patients with moderate to severe melasma. It would also
This article is protected by copyright. All rights reserved.

need to be modified for use in patients with excess facial hair and concomitant pigmentary
disorders.

In conclusion, we have developed and validated a novel algorithm for automated scoring of aMASI
that correlates closely with dermatologist scored mMASI. aMASI has also been shown to be sensitive
enough to pick up changes in MASI in response to treatment. This aMASI scoring represents a
promising, novel and reliable method for the assessment of disease severity and treatment response
to melasma.

References

1. Grimes PE. Melasma: Etiologic and therapeutic considerations. Arch Dermatol 1995; 131:
1453-7
2. Sheth VM, Pandya AG. Melasma: A comprehensive update. J Am Acad Dermatol 2011; 65:
689-97
3. Tamega AA, Miot LDB, Bonfietti C et al. Clinical patterns and epidemiological characteristics
of facial melasma in Brazilian women. J Eur Acad Dermatol Venereol 2013; 27: 151-6
4. Handel AC, Lima PB, Tonolli VM et al. Risk factors for facial melasma in women: a case-
control study. Br J Dermatol. Advance online publication. doi: 10.1111/bjd.13059
5. Balakrishnan R, McMichael AJ, Camacho FT et al. Development and validation of a health-
related quality of life instrument for women with melasma. Br J Dermatol 2003; 149: 572-77
6. Kimbrough-Green CK, Griffiths CEM, Finkel LJ et al. Topical retinoic acid (tretinoin) for
melasma in black patients. Arch Dermatol 1994; 130: 727-33
7. Pandya AG, Hynan LS, Bhore R, Riley FC et al. Reliability assessment and validation of the
Melasma Area and Severity Index (MASI) and a new modified MASI scoring method. J Am
Acad Dermatol 2011; 64: 78-83
8. Khan R, Hanbury A, Stöttinger J et al. Colour based skin classification. Pattern Recognition
Letters, 2012; 33: 157-63
9. Ramsay B, Lawrence CM. Measurement of involved surface area in patients with psoriasis.
Br J Dermatol 1991; 124: 565-70
10. Puzenat E, Bronsard V, Prey S et al. What are the best outcome measures for assessing
plaque psoriasis severity? A systematic review of the literature. J Eur Acad Dermatol
Venereol 2010; 24: 10-6
11. Espinoza E, Martinez G, Frerichs JF et al. Cell cluster segmentation based on global and local
thresholding for in-situ microscopy. Biomedical Imaging: Nano to Macro; 2006. 3rd IEEE
International symposium 2006, 542-5
12. Zemouri E, Chibani Y, Brik Y. Enhancement of Historical Document Images by Combining
Global and Local Binarization Technique. International Journal of Information and Electronics
Engineering 2014; 4: 1-5
13. Gabarra E, Tabbone A. Combining Global and Local Threshold to Binarize Document of
Images. Pattern Recgonition and Image Analysis 2005; 3523: 371-8
This article is protected by copyright. All rights reserved.

Legends

Fig. 1: aMASI score was calculated based on the mMASI formula as shown above. A, Percentage area
involved; D, darkness, L, left, R, right.

Fig.2: (a) The pixels around the eyes, eyebrows, nostrils and mouth automatically have a larger
difference in colour compared to the reference point of normal skin. This difference is usually above
a certain threshold. By colour thresholding; pixels around the eyes, eyebrows, nostrils and mouth
are classified into the non-skin set. (b) The computer can then automatically identify skin (white) and
non-skin (black) areas. Non-skin areas would automatically be excluded from the analysis.

Fig.3: (a) The area of the chin is defined by the area beneath a curve that passes through the corners
of the mouth and the centre of the nostrils. (b) The left malar region is determined from the left
lateral view. It is bounded by a line drawn from the left eyelid to the frontal hairline superiorly, a
curve from the nostril to the left angle of mouth inferiorly and the jawline laterally.

Fig.4: (a) Melasma before segmentation by computer. (b) Melasma area as segmented by the
algorithm for analysis and mMASI calculation.

Fig.5: Scatter plot showing relationship between aMASI and clinical mMASI at baseline.

Fig.6: Scatter plot showing relationship between aMASI and clinical mMASI post-treatment

Fig.7: Scatter plot showing relationship between change in aMASI and change in clinical mMASI pre-
and post-treatment
This article is protected by copyright. All rights reserved.

Table 1: Comparison of clinician scored mMASI scores and automated mMASI scores

Clinical mMASI (n=29) Automated mMASI (n=29)

Baseline Post Change in Baseline Post Change in

Treatment mMASI Treatment mMASI

Mean 3.64 3.16 -0.48 3.54 3.33 -0.21

Standard 2.45 2.26 0.87 1.31 1.42 0.67

deviation

Minimum 0.83 0.63 -2.40 1.96 1.62 -1.72

Maximum 9.90 10.8 1.10 8.05 9.50 1.45

This article is protected by copyright. All rights reserved.
This article is protected by copyright. All rights reserved.
This article is protected by copyright. All rights reserved.
This article is protected by copyright. All rights reserved.