cosmetics

Article
The Development of a Skin Image Analysis Tool by
Using Machine Learning Algorithms
Perry Xiao 1, * , Xu Zhang 2 , Wei Pan 3 , Xiang Ou 4 , Christos Bontozoglou 1 , Elena Chirikhina 1
and Daqing Chen 1
1 School of Engineering, London South Bank University, 103 Borough Road, London SE1 0AA, UK;
bontozoc@lsbu.ac.uk (C.B.); chirikhe@lsbu.ac.uk (E.C.); chend@lsbu.ac.uk (D.C.)
2 Auckland Tongji Medical & Rehabilitation Equipment Research Centre, Tongjing Zhejiang College,
JiaXing 314000, China; zhangx12@tjzj.edu.cn
3 School of Computer Science and Engineering, Changshu Institute of Technology, Changshu 215500, China;
pvv1224@163.com
4 College of Communication Engineering, Chongqing University, Chongqing 400044, China;
ouxiang@cqu.edu.cn
* Correspondence: xiaop@lsbu.ac.uk; Tel.: +44-20-7815-7569; Fax: +44-20-7815-7561




Received: 23 July 2020; Accepted: 20 August 2020; Published: 22 August 2020


Abstract: We present our latest research work on the development of a skin image analysis tool by
using machine-learning algorithms. Skin imaging is very import in skin research. Over the years,
we have used and developed different types of skin imaging techniques. As the number of skin
images and the type of skin images increase, there is a need of a dedicated skin image analysis tool.
In this paper, we report the development of such software tool by using the latest MATLAB App
Designer. It is simple, user friendly and yet powerful. We intend to make it available on GitHub,
so that others can benefit from the software. This is an ongoing project; we are reporting here what
we have achieved so far, and more functions will be added to the software in the future.

Keywords: graphic user interface; deep learning; machine learning; capacitive imaging; skin
image analysis

1. Introduction
Skin imaging is very import in skin research. To date, there are many popular skin imaging
technologies. Dermatoscopy is probably the most widely used skin imaging technology and is
commonly used to examine skin lesions with a dermatoscope [1,2]. A dermatoscope typically consists
of a magnifier, a light source (polarized or non-polarized) and a transparent plate. Skin photos and
videos can be recorded by using a digital camera. DermLite is a small dermatoscope that can be
attached to a smart phone. It claims to be the world’s best-selling dermatoscope [3]. DermLite II
MultiSpectral is another pocket size dermatoscope, it is button activated, with cross-polarized 4-color
imaging [4]. Nerasolutions provides several imaging devices, not only for skin imaging, but also
for hair imaging [5]. Visia Skin Analysis System takes multiple photos of face and analyzes the
melanin distributions, the texture, the wrinkle, the pores, etc. [6]. EPISCAN I-200 is a high resolution
ultrasound imaging system that can take skin images using ultrasound up to 50 MHz [7]. Last not
the least, Epsilon permittivity imaging system is a capacitive image technology developed in our
research group [8–13]. It is based on dielectric constant measurement principle, and has been used for
skin hydration imaging, skin solvent penetration imaging, as well as hair and nail imaging. With all
these skin imaging technologies, there is a genuine need for a general purpose, user friendly and yet
powerful skin software analysis tool. In this paper, we present our latest work on the development of a

Cosmetics 2020, 7, 67; doi:10.3390/cosmetics7030067 www.mdpi.com/journal/cosmetics

Cosmetics 2020, 7, 67 2 of 12
Cosmetics 2020, 7, x FOR PEER REVIEW 2 of 13

development of a skin
skin image analysis tool image analysis
based on tool based on
machine-learning machine-learning
algorithms. algorithms.
We will first We
present the will first
design and
present the design
the structure and the
of the skin structure
software of the
analysis skin
tool, software
then presentanalysis tool,
the results then
and present the
discussions andresults
finally,and
the
discussions
future works.and finally, the future works.

2. Materials
Materials and
and Methods

2.1. Front-End
2.1. Front-End GUI
GUI
The front-end
The front-end GUIGUI (graphical
(graphical user
user interface)
interface) is
is designed
designed and and developed
developed by by using
using MATLAB
MATLAB
App
App Designer,
Designer, which
which is is relatively
relatively new
new and
and provides
provides much
much better
better user
user interface
interface components
components and and aa
better integrated development environment than the previous MATLAB
better integrated development environment than the previous MATLAB GUIDE. With MATLAB GUIDE. With MATLAB App
Designer, you can easily design and develop a user friendly software interface. It
App Designer, you can easily design and develop a user friendly software interface. It offers a grid offers a grid layout
manager
layout to organize
manager the user
to organize theinterface and automatic
user interface reflowreflow
and automatic options to make
options the app
to make the detect and
app detect
respond to changes in screen size. The coding is also a lot simpler, with machine
and respond to changes in screen size. The coding is also a lot simpler, with machine generated code generated code
automatically grayed
automatically grayed out.
out. Figure
Figure 11 shows
shows the
the GUI
GUI ofof the
the skin
skin image
image analysis
analysis software
software tool. The left
tool. The left
panel shows
panel shows the
the original
original image
image andand analyzed
analyzed results. The right
results. The right panel
panel contains
contains several
several tabs
tabs with
with each
each
tab is dedicated to different functions. Apart from reading from image files, it
tab is dedicated to different functions. Apart from reading from image files, it can also get image can also get image
frames from
frames from webcams
webcams and and other
other USB
USB based
based video
video devices.
devices. This
This is
is done
done byby using
using “MATLAB
“MATLAB Support
Support
Package for USB Webcams” Add-On module which has shown significant
Package for USB Webcams” Add-On module which has shown significant performance performance improvement
than the existing
improvement thanvideo input function
the existing from function
video input the imagefromacquisition
the imagetoolbox.
acquisition toolbox.

Figure
Figure 1. The
The GUI (graphical user interface) of the skin image analysis tool.

2.2. Back-End
2.2. Back-End MATLAB
MATLAB Files
Files
The back-end
The back-end of of the
the program
program is is implemented
implemented inin aa number
number ofof MATLAB
MATLAB functions saved as
functions saved as
standard M-files. MATLAB functions, also called modules or subroutines in other
standard M-files. MATLAB functions, also called modules or subroutines in other programming programming
languages, accept
languages, accept input
input arguments
arguments andand return
return output
output arguments. They operate
arguments. They operate on
on variables within
variables within
their own
their ownworkspace,
workspace,different from
different thethe
from main program.
main TableTable
program. 1 shows the structure
1 shows of the skin
the structure analysis
of the skin
software tool.
analysis software tool.
Table 1. Structure of the skin analysis tool.
Table 1. Structure of the skin analysis tool.
Files Comments
Files Comments
PX_DL_GUI.mlapp
PX_DL_GUI.mlapp Main GUI
Main GUI filefile
AlexnetTraining.m
AlexnetTraining.m AlexNet
AlexNettraining
trainingfunction
function
AlexnetClassify.m AlexNet classification function
AlexnetClassify.m AlexNet classification function
GoogLeNetTraining.m GoogLeNet training function
GoogLeNetTraining.m GoogLeNet training function
GoogLeNetClassify.m GoogLeNet classification function
GoogLeNetClassify.m
Gabor_Calculate.m GoogLeNet
Gabor classification
wavelet function function
transform calculation
Gabor_Calculate.m
Gabor_Search.m Gabor wavelet transform
Gabor wavelet transform search calculation function
function
Gabor_Search.m
... ... .Gabor
. . . . . wavelet transform search function
…… ……
Cosmetics 2020, 7, 67 3 of 12

2.3. Functionalities
Cosmetics 2020, 7, x FOR PEER REVIEW 3 of 13

The main purpose of this work is to develop a general purpose skin image analysis software with
2.3. Functionalities
many useful skin image analysis functions. It is also designed to be flexible, so more new functions can
The main purpose of this work is to develop a general purpose skin image analysis software
be added in the future.
with many useful skin image analysis functions. It is also designed to be flexible, so more new
The current
functions canimplemented
be added in thefunctions
future. are image classifications [8,9,13], skin texture analysis and
skin image retrieval by using Gabor wavelet transform,
The current implemented functions are image PCA (principle
classifications components
[8,9,13], skin analysis)
texture analysis and and
GLCM (gray-level
skin co-occurrence
image retrieval matrix)
by using Gabor [10,11],
wavelet as well as
transform, PCAskin(principle
live image analysis.analysis) and
components
GLCM (gray-level co-occurrence matrix) [10,11], as well as skin live image analysis.
3. Results and Discussions
3. Results and Discussions
3.1. Skin Image Classification
3.1. Skin Image Classification
Skin image classification is the most important function of the skin analysis tool. With skin image
Skin we
classification, imagewouldclassification
be able to is the most important
differentiate normalfunction of the
skin from skin analysis
damaged skin,tool. Withskin
healthy skinfrom
image classification, we would be able to differentiate normal skin from
diseased skin. It can also potentially be used for identifying different types of skin diseases damaged skin, healthy skin and
from diseased skin. It can also potentially be used for identifying different types of skin diseases and
skin cancers. Skin image classification is achieved through transfer learning by using pre-trained
skin cancers. Skin image classification is achieved through transfer learning by using pre-trained
deep-learning neural networks such as AlexNet [14], GoogLeNet [15], VGG19 [16] and ResNet101 [17].
deep-learning neural networks such as AlexNet [14], GoogLeNet [15], VGG19 [16] and ResNet101
The performance of image
[17]. The performance classification
of image is tested
classification ononfour
is tested fourdifferent typesofofimages,
different types images, as shown
as shown in in
FiguresFigures 2–5. All the images are randomly divided into training set (70%) and test set (30%) and and
2–5. All the images are randomly divided into training set (70%) and test set (30%)
automatically resized
automatically to 224
resized × 224
to 224 × 224××33(except
(except for AlexNet,227
for AlexNet, × 227
227× 227 × 3)×before
3) before training.
training.
The grayscale
The grayscale images
imagesalso automatically
also automaticallyconverted
converted to tocolor
colorimages
images prior
prior to training.
to training. During
During the the
training, it uses stochastic gradient descent with momentum, the initial learning
training, it uses stochastic gradient descent with momentum, the initial learning rate is 10ˆ-4, mini batch rate is 10^-4, mini
size isbatch sizeepochs
5, max is 5, max epochs
is 6, and itis also
6, and it also shuffles
shuffles trainingtraining
between between every epoch.
every epoch. The trainings
The trainings are on
are done
done on a standard desktop computer ®with Intel® Core™ i7-3770 CPU @3.4 GHz, 8 cores, 16 GB
a standard desktop computer with Intel Core™ i7-3770 CPU @3.4 GHz, 8 cores, 16 GB RAM and
RAM and Windows 8.1 operating system.
Windows 8.1 operating system.

palm neck leg

face forearm palm

Figure
Figure 2. Sample
2. Sample skinskin capacitive
capacitive images
images from
from different
different skin
skin sitesofofa ahealthy
sites healthyvolunteer.
volunteer. There
There are
are total
total 90
90 images images into
divided divided into six categories,
six categories, palm, palm,
neck, neck, leg, face,
leg, face, forearm
forearm andand palm.
palm.

Figure
Figure 2 shows
2 shows thethe
skinskin capacitiveimages
capacitive images acquired
acquired bybyusing
usingour Epsilon
our Epsilonpermittivity imaging
permittivity imaging
system [8,9]. There are total 90 images divided into six categories, palm, neck, leg, face, forearm and
system [8,9]. There are total 90 images divided into six categories, palm, neck, leg, face, forearm and
palm. The image classification results are shown in the first row of Table 2. AlexNet has the shortest
palm. The image classification results are shown in the first row of Table 2. AlexNet has the shortest
training time and VGG19 has the longest training time. The classification overall accuracy is good,
except VGG19 and GoogLeNet gives an impressive 100% classification accuracy.
Cosmetics 2020, 7, x FOR PEER REVIEW 4 of 13

training2020,
Cosmetics time7, 67
and
VGG19 has the longest training time. The classification overall accuracy is good,
4 of 12
except VGG19 and GoogLeNet gives an impressive 100% classification accuracy.

Benign

Malignant

2020,3.7,Sample
Figure
Cosmetics skin REVIEW
x FOR PEER cancer images from ISIC database [18]. There are total 600 images divided into5 of 13
two categories,
Figure 3. SampleBenign and Malignant.
skin cancer images from ISIC database [18]. There are total 600 images divided into
two categories, Benign and Malignant.

Figure 3 shows the skin cancer images from ISIC database [18]. These are standard digital photo
images. There are totally 600 images divided into two categories: Benign and Malignant. The image
classification results are shown in the second row of Table 2. AlexNet again has the shortest training
time, and VGG19 has the longest. GoogLeNet and ResNet101 have the best classification accuracy.
Figure 4 shows the skin ultrasound images acquired by using EPISCAN I-200 ultrasound
instrument. There are totally 610 images measured from six different skin sites, chin, cheeks, nose,
under eye, lips and forearm. The image classification results are shown in the third row of Table 2.
AlexNet again has the shortest training time and VGG19 the longest. ResNet101 has the best
classification accuracy. The accuracy is slightly lower than lower than previous two types of skin
images, this is understandable, as skin ultrasound images have higher similarities.

Chin cheeks nose

under eye lips forearm

4. Sample skin ultrasound images from different skin sites of aa healthy

Figure 4. healthy volunteer.
volunteer. There are
into six
total 610 images divided into six categories,
categories, chin,
chin, cheeks,
cheeks, nose,
nose, under
under eye,
eye, lips
lips and
and forearm.
forearm.
Cosmetics 2020, 7, 67 5 of 12

Figure 3 shows the skin cancer images from ISIC database [18]. These are standard digital photo
images. There are totally 600 images divided into two categories: Benign and Malignant. The image
classification results are shown in the second row of Table 2. AlexNet again has the shortest training
time, and VGG19 has the longest. GoogLeNet and ResNet101 have the best classification accuracy.
Figure 4 shows the skin ultrasound images acquired by using EPISCAN I-200 ultrasound
instrument. There are totally 610 images measured from six different skin sites, chin, cheeks, nose,
under eye, lips and forearm. The image classification results are shown in the third row of Table 2.
AlexNet again has the shortest training time and VGG19 the longest. ResNet101 has the best
classification accuracy. The accuracy is slightly lower than lower than previous two types of skin
images,Cosmetics
this is 2020, 7, x FOR PEER REVIEW
understandable, as skin ultrasound images have higher similarities. 6 of 13

COVID-19

Normal

Figure
Figure 5. 5. Chest
Chest X-rayX-ray images
images from
from University of
University of Montreal
Montreal [19] and
[19] Kaggle
and [20]. [20].
Kaggle ThereThere
are total
are50total 50
images divided into two categories, COVID-19 and normal.
images divided into two categories, COVID-19 and normal.
Table 2. Image classification results.
Currently, the coronavirus COVID-19 has become a global pandemic and is affecting more than
21 million people in more Table Column Head
Imagethan
Types200 countries and territories around the world. We wanted to see
if our software tool can effectively work on COVID-19 chest X-rayAccuracy
Models Training Time
images, as shown in Figure 5.
AlexNet 1 min 6 s 83.33%
The COVID-19 chest X-ray images are from a research group in University of Montreal [19] and
Capacitive GoogLeNet 1 min 39 s 100.00%
the normal chest X-rayskinimages
imagesare fromVGG19
Kaggle [20]. 12There
min 5are
s totally 50 images divided into two
58.33%
categories: COVID-19 and normal. The image classification
ResNet101 9 min 12 sresults 87.50%
are shown in the third row of
Table 2. AlexNet, not surprisingly, has the shortest training
AlexNet time
8 min 37 s and VGG19
73.89% the longest. The overall
Skin cancer GoogLeNet 13 min 23
classification accuracy is high (>85%) and ResNet101 has an impressive s 77.78%
100% classification accuracy.
images VGG19 91 min 15 s 75.00%
In summary, AlexNet is consistently the fastest in image classification training, while VGG19
ResNet101 66 min 40 s 77.78%
is the slowest. The ResNet101 has the highest overall classification accuracy and VGG19 has the
AlexNet 4 min 56 s 69.44%
lowest overall classificationSkin
accuracy. For skin capacitive image classification, GoogLeNet is the best.
GoogLeNet 10 min 51 s 56.67%
ultrasound
For other image classification, e.g., skin cancer
VGG19images, 71skin ultrasound
min 59 s images
60.56% and Chest X-ray images,
images
ResNet101 is the best. ResNet101 63 min 3 s 70.56%
AlexNet 43 s 92.86%
Chest X-ray GoogLeNet 1 min 30 s 92.85%
images VGG19 6 min 44 s 85.71%
ResNet101 4 min 42 s 100.00%

Currently, the coronavirus COVID-19 has become a global pandemic and is affecting more than
21 million people in more than 200 countries and territories around the world. We wanted to see if
our software tool can effectively work on COVID-19 chest X-ray images, as shown in Figure 5. The
Cosmetics 2020, 7, 67 6 of 12

Table 2. Image classification results.

Table Column Head

Image Types
Models Training Time Accuracy
AlexNet 1 min 6 s 83.33%
GoogLeNet 1 min 39 s 100.00%
Capacitive skin images
VGG19 12 min 5 s 58.33%
ResNet101 9 min 12 s 87.50%
AlexNet 8 min 37 s 73.89%
GoogLeNet 13 min 23 s 77.78%
Skin cancer images
VGG19 91 min 15 s 75.00%
ResNet101 66 min 40 s 77.78%
AlexNet 4 min 56 s 69.44%
GoogLeNet 10 min 51 s 56.67%
Skin ultrasound images
VGG19 71 min 59 s 60.56%
ResNet101 63 min 3 s 70.56%
AlexNet 43 s 92.86%
GoogLeNet 1 min 30 s 92.85%
Chest X-ray images
VGG19 6 min 44 s 85.71%
ResNet101 4 min 42 s 100.00%

The most significant benefit of skin image classification is that users would be able to use the skin
image analysis tool with a webcam camera to take a picture of a skin lesion, the software will be able
to tell whether it is benign or malignant. Users can also train the neural networks on their own skin
image data, to recognize different types of skin diseases/cancers. This could be a potential low cost
skin diseases/cancer diagnosis device.
The classification accuracy depends on the number and the quality of skin images, as well as
the deep-learning neural networks. To further improve the accuracy, we will need train the neural
networks on larger number of skin images and better quality of skin images. Currently we only have
tested four different deep-learning neural networks, in the future, we will implement more neural
networks. In addition, all the neural networks are trained on a set of default optimized values, next step
is to allow users to specify their own training parameter values.

3.2. Skin Texture Analysis and Image Retrieval

Skin surface is not even, it is rich with patterns and lines. Our previous studies show that we can
analyze the skin texture by using Gabor wavelet transform, PCA (principle components analysis) and
GLCM (gray-level co-occurrence matrix). With Gabor wavelet transform, we can get skin features
vectors on color, shape and texture [10]. With GLCM (gray-level co-occurrence matrix), we can get
four GLCM feature vectors, e.g., angular second moment (ASM), entropy (ENT), contrast (CON) and
correlation (COR). The results showed that the angular second moment increases as age increases and
entropy decreases as age increases [11]. We can also analyze the skin micro-relief, the results showed
that the change in the intensity of primary and secondary lines during arm extension is increasing
against chronological age, while the average number of closed polygons per square mm is decreasing
against chronological age [12].
Another useful usage of skin texture analysis is image retrieval or image searching. This is
particularly suitable for skin images, where similarities are much higher than normal images. Figure 6a
shows the skin cancer image searching results using Gabor wavelet transform. As we use the same
image database for training and searching, the best matching result is the query image itself. The second
to sixth best matching results are also from the same category of the query image, i.e., benign images.
Figure 6b–d show the image searching results using Gabor wavelet transform on skin capacitive
images, skin ultrasound images and chest X-ray images. Similar to Figure 6a, all the best matching
results are from the same category of the query image.
Cosmetics 2020, 7, 67 7 of 12

Table 3 shows the image searching results of different algorithms on different types of images.
Calculation time is how long each algorithm takes to calculate the feature vectors of all the images
and built a database of the results. Searching time is how long each algorithm takes to search the
database and find the six best matching results. The error rate is calculated as how many images in the
best six matching results are in wrong category. Figure 6 shows the best scenario of Gabor wavelet
transform searching results, the error rates are 0 out of 6, while Table 3 shows the average error rates.
The overall results show that Gabor wavelet transform general takes longer time to calculate and to
search, while PCA is the shortest for calculation and GLCM is the shortest for searching. Although
Gabor wavelet transform is slower compared with other image retrieval techniques, such as PCA and
GLCM, Gabor wavelet transform is the best for retrieving images. This agrees well our previous studies
on skin capacitive contact images and digital facial images [10]. All the algorithms are running on a set
Cosmetics 2020, 7, x FOR PEER REVIEW 8 of 13
of optimized default values, in the future, we will allow users to specify their own parameter values.

(a)

(b)

Figure 6. Cont.
Cosmetics 2020,
Cosmetics 2020, 7,
7, 67
x FOR PEER REVIEW 98 of
of 13
12

(c)

(d)

Figure 6. Image searching results using Gabor wavelet transform on different type of images. (a) skin
Figure 6. Image
capacitive searching
images; (b) skinresults
cancer using Gabor
images; wavelet
(c) skin transform
ultrasound on different
images; (d) chesttype of images.
X-ray images. (a) skin
capacitive images; (b) skin cancer images; (c) skin ultrasound images; (d) chest X-ray images.
Table 3. Image searching results.
Table 3 shows the image searching resultsTableof different algorithms on different types of images.
Column Head
Image Types
Calculation time is how long each algorithm takes to calculate the feature vectors of all the images
Algorithms Calculation Time Searching Time Error Rate
and built a database of the results. Searching time is how long each algorithm takes to search the
Gabor results. The
database and find the six best matching 153 s
error 1059 msas how many
rate is calculated 1/6 images in
Capacitive skin images PCA 1.6 s 221 ms 3/6
the best six matching results are in wrong category. Figure 6 shows the best scenario of Gabor
GLCM 1.1 s 25 ms 2/6
wavelet transform searching results, the error rates are 0 out of 6, while Table 3 shows the average
Gabor 370 s 2834 ms 0/6
error rates.
SkinThe overall
cancer results show that Gabor wavelet transform general takes longer time to
images PCA 54 s 1560 ms 2/6
calculate and to search, while PCA GLCM is the shortest
482for
s calculation and GLCM is the2/6
698 ms shortest for
searching. Although Gabor wavelet transform is slower compared with other image retrieval
techniques, such as PCA and GLCM, Gabor wavelet transform is the best for retrieving images. This
agrees well our previous studies on skin capacitive contact images and digital facial images [10]. All
Cosmetics 2020, 7, x FOR PEER REVIEW 10 of 13

the algorithms are running on a set of optimized default values, in the future, we will allow users to
specify their own parameter values.
Cosmetics 2020, 7, 67 9 of 12
Table 3. Image searching results.

Table Table
Column Head
3. Cont.
Image Types Calculation
Algorithms Searching Time Error Rate
Table Column Head
Time
Image Types
Gabor Algorithms 153 s
Calculation Time 1059 ms
Searching Time 1/6Rate
Error
Capacitive
PCA Gabor 1.6 s 324 s 221 ms2743 ms 3/6
2/6
skin images
GLCM PCA
Skin ultrasound images 1.1 s 29 s 25 ms2409 ms 2/6
3/6
Gabor GLCM 370 s 118 s 2834 ms220 ms 1/6
0/6
Skin cancer
PCA Gabor 54 s 35 s 1560 ms 813 ms 2/6
0/6
images
Chest X-ray images
GLCM PCA 482 s 2.6 s 698 ms 162 ms 2/6
2/6
GLCM 30 s 563 ms 0/6
Skin Gabor 324 s 2743 ms 2/6
ultrasound PCA 29 s 2409 ms 3/6
3.3. Skin Live Image Analysis
images GLCM 118 s 220 ms 1/6
from analyzeGabor
Apart X-ray
Chest
35 s analyze skin live
image files, it can also 813 images
ms 0/6 webcams
captured from
and other PCA 2.6 s 162 ms
USB based video devices. The live image capture is achieved by using the latest 2/6 MATLAB
images
GLCM Support Package
Add-On module called “MATLAB 30 s for USB Webcams”.
563 ms Figure 7 shows 0/6the live image
analysis of facial skin using the ProScope HR with 50 times magnification [21]. The facial features such
3.3. Skintexture,
as skin Live Image Analysis
speckles and facial hair are clear to see. The software can divide the live image into
threeApart
color channels,
from analyze green
red, imageand blue
files, it (RGB)
can also and therefore
analyze skinperform the RGB
live images calculation
captured frominwebcams
real time.
Figure
and 7a shows
other USB basedthe original image and
video devices. The processed
live image image (green
capture channel
is achieved byonly),
usingFigure 7b shows
the latest MATLAB the
original image and processed image (red channel only). As we can see,
Add-On module called “MATLAB Support Package for USB Webcams”. Figure 7 shows the live green channel has more surface
information,
image analysis while red channel
of facial has more
skin using information
the ProScope HRonwith
melanin (facial
50 times hair) and skin[21].
magnification color.
The facial
features such as skin texture, speckles and facial hair are clear to see. The software can divide channel
Figure 7c shows the original image and the processed image, calculated by adding green the live
× 0.7 with
image intoblue
three color ×channels,
channel 0.7, without
red,red channel.
green and As we (RGB)
blue can see,and
this therefore
enhances the surface
perform thetexture
RGB
informationin
calculation of the
realskin.
time.Users
Figurecan7a
specify
shows they
theown RGB channel
original image andcalculations.
processed Facial
imageskin(green
has enormous
channel
cosmetic
only), values,
Figure 7b this
showsfunctionality
the originalcould
imagebe very useful for analyzing
and processed image (red skin facial features,
channel only). Assuch as facial
we can see,
hairs, smoothness, speckles, scars, wrinkles, all in real time! More functions
green channel has more surface information, while red channel has more information on melaninwill be added. The current
versionhair)
(facial of the
andprogram is available as a MATLAB App Installer format on GitHub [22].
skin color.

(a)

Figure 7. Cont.
Cosmetics 2020, 7,
Cosmetics 2020, 7, 67
x FOR PEER REVIEW 10
11 of
of 12
13

(b)

(c)

Figure 7.
Figure 7. Live
Live image
image analysis
analysis using
using the
the ProScope
ProScope HR HR with
with 50
50 times
times magnification.
magnification. (a)
(a) original
original image
image
and processed
and processedimage
image(green
(greenchannel
channel only);
only); (b) (b) original
original image
image and and processed
processed imageimage (red channel
(red channel only);
only); (c) original image and the processed image (no red channel, green × 0.7 +
(c) original image and the processed image (no red channel, green × 0.7 + blue × 0.7). blue × 0.7).

4. Conclusions
Figure 7c shows the original image and the processed image, calculated by adding green
channel × 0.7 developed
We have with blue channel × 0.7, analysis
a skin image without tool
red based
channel.
on As we can see, thisalgorithms.
machine-learning enhances the surface
It uses the
texture information of the skin. Users can specify they own RGB channel calculations.
latest MATLAB App Designer to create the graphic user interface on the front end and use MATLAB Facial skin has
enormous
M-files cosmetic
as the values,
backend. Thethis functionality
current version could be very
provides usefulsuch
functions for analyzing skin facial
as skin image features,
classifications,
such image
skin as facial hairs, smoothness,
retrieval and skin livespeckles, scars, wrinkles,
image analysis. The skinallimage
in realclassification
time! More isfunctions
done bywill be
using
added. The current version of the program is available as a MATLAB App Installer
deep-learning neural networks such as AlexNet, GoogLeNet, VGG19 and ResNet101. The classification format on
GitHub [22].
Cosmetics 2020, 7, 67 11 of 12

results show that GoogLeNet is the best for skin capacitive image classifications, while ResNet101 is the
best for skin cancer image, skin ultrasound image and chest X-ray image classifications. The biggest
advantage of this skin image analysis tool is that users would be able to retrain the deep-learning neural
networks on their own skin image data and therefore perform their own skin image classifications,
such as recognizing different types of skin diseases. This could lead to a potential low cost skin
disease/cancer diagnose device. The skin image retrieval is done by using Gabor wavelet transform,
PCA and GLCM. The results show that Gabor wavelet transform has the overall best accuracy despite
of being the slowest. In skin live image analysis, by manipulating the RGB values of the images,
it has potentials for analyzing skin facial features, such as facial hairs, smoothness, speckles, scars,
wrinkles in real time. More functionality will be added on in the future. The software tool can work on
a variety of different types of skin images, such as skin digital images, skin capacitive images, skin
ultrasound images and more. Our study shows the apart from skin images, it can also work on other
types of images, such as chest X-ray images. We have made the program available on GitHub so that
others can benefit from it.
For the future work, we would like to test the skin image analysis tool on larger skin image
dataset, on more volunteers, of different ages, genders, races and colors, as well as different skin
conditions (dry, normal, silky, rough, etc.) and different skin diseases/cancers. For image classifications,
more deep-learning neural networks will be implemented, and users would be able to select different
training parameters. For skin image retrieval, users would be able to specify different parameter values
for Gabor wavelet transform, PCA and GLCM. For live image analysis, more functions, such as image
segmentation, object detections and so on, will be implemented.

Author Contributions: Conceptualization, P.X. and D.C.; methodology, P.X., X.O., X.Z., W.P., C.B. and E.C.; formal
analysis, P.X., X.O., X.Z., W.P. and C.B.; investigation, X.O., X.Z., W.P. and E.C.; data curation, P.X. and C.B.; original
draft preparation, P.X.; review and editing of manuscript, P.X. and D.C.; supervision, P.X. and D.C. All authors
have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: We thank the London South Bank University and Biox Systems, Ltd., UK for the financial
support. We also thank Longport, Inc., USA for loaning the EPISCAN I-200 instrument.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Argenziano, G.; Soyer, H.P. Dermoscopy of pigmented skin lesions–a valuable tool for early diagnosis of
melanoma. Lancet Oncol. 2011, 2, 443–449. [CrossRef]
2. Vestergaard, M.E.; Macaskill, P.; Holt, P.E.; Menzies, S.W. Dermoscopy compared with naked eye examination for
the diagnosis of primary melanoma: A meta-analysis of studies performed in a clinical setting. Br. J. Dermatol.
2008, 159, 669–676. [CrossRef] [PubMed]
3. Dermlite. Available online: https://dermlite.com/ (accessed on 18 March 2020).
4. DermLite II Multi-Spectral. Available online: https://www.eufoton.com/en/products/accessories/dermlite-
multispectral-ii (accessed on 18 March 2020).
5. Nerasolutions. Available online: https://www.nerasolutions.co.uk/skin-analysis-devices (accessed on
18 March 2020).
6. Visia Skin Analysis System. Available online: https://www.canfieldsci.com/imaging-systems/visia-
complexion-analysis/ (accessed on 18 March 2020).
7. EPISCAN I-200. Available online: https://www.longportinc.com/about-longport (accessed on 18 March 2020).
8. Zhang, X.; Pan, W.; Bontozoglou, C.; Chirikhina, E.; Chen, D.; Xiao, P. Skin Capacitive Imaging Analysis
Using Deep Learning GoogLeNet. In Intelligent Computing. SAI 2020. Advances in Intelligent Systems and
Computing; Arai, K., Kapoor, S., Bhatia, R., Eds.; Springer: Cham, Switzerland, 2020; Volume 1229. [CrossRef]
9. Zhang, X.; Pan, W.; Xiao, P. In-vivo Skin Capacitive Image Classification Using AlexNet Convolution Neural
Network. In Proceedings of the 2018 IEEE 3rd International Conference on Image, Vision and Computing
(ICIVC), Chongqing, China, 27–29 June 2018; pp. 439–443. [CrossRef]
Cosmetics 2020, 7, 67 12 of 12

10. Ou, X.; Pan, W.; Zhang, X.; Xiao, P. Skin image retrieval using Gabor wavelet texture feature. Int. J. Cosmet.
Sci. 2016, 38, 607–614. [CrossRef] [PubMed]
11. Ou, X.; Pan, W.; Xiao, P. In-Vivo Skin Capacitive imaging Analysis by using Grey Level Co-occurrence Matrix
(GLCM). Int. J. Pharm. 2014, 460, 28–32. [CrossRef] [PubMed]
12. Bontozoglou, C.; Zhang, X.; Xiao, P. Micro-relief analysis with skin capacitive imaging. Skin Res. Technol.
2019, 25, 165–170. [CrossRef] [PubMed]
13. Bianconi, F.; Chirikhina, E.; Smeraldi, F.; Bontozoglou, C.; Xiao, P. Personal identification based on skin
texture features from the forearm and multi-modal imaging. Skin Res. Technol. 2017, 23, 392–398. [CrossRef]
[PubMed]
14. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet classification with deep convolutional neural networks.
In Proceedings of the International Conference on Neural Information Processing Systems, Lake Tahoe, NV,
USA, 3–8 December 2012; pp. 1097–1105.
15. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A.
Going Deeper With Convolutions. In Proceedings of the The IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Boston, MA, USA, 7–12 June 2015; pp. 1–9.
16. Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition.
ICLR 2015, arXiv:1409.1556v6.
17. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. arXiv 2015, arXiv:1512.03385.
18. ISIC—The International Skin Imaging Collaboration (Skin Disease Database). Available online: https:
//isic-archive.com/#images (accessed on 20 March 2020).
19. University of Montreal’s X-ray Images of COVID-19 Cases. Available online: https://github.com/ieee8023/
covid-chestxray-dataset (accessed on 20 March 2020).
20. Kaggle Chest X-Ray Images (Pneumonia). Available online: https://www.kaggle.com/paultimothymooney/
chest-xray-pneumonia (accessed on 20 March 2020).
21. ProScope, H.R. The ProScope Mobile & HR USB Handheld Digital Microscope. Available online: https:
//www.bodelin.com/ (accessed on 20 July 2020).
22. Skin Image Analysis Tool GitHub Site. Available online: https://github.com/PerryXiao2015/LSBU-Skin-
Image-Analysis-Tool (accessed on 21 August 2020).

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).