Accelerat ing t he world's research.

Deep Learning Application for

Effective Classification of Different
Types of Psoriasis
Fahad Azim
Journal of Healthcare Engineering

Cite this paper Downloaded from Academia.edu 

Get the citation in MLA, APA, or Chicago styles

Related papers Download a PDF Pack of t he best relat ed papers 

A Deep Learning Approach t o Det ect COVID-19 Pat ient s from Chest X-ray Images
Khandaker Foysal Haque

Design and Development of a Humanoid Robot for Sign Language Int erpret at ion
Sejut i Rahman

Art ificial Neural Net work-Based Deep Learning Model for COVID-19 Pat ient Det ect ion Using X-Ray Che…
Mohammad Shorfuzzaman
Hindawi
Journal of Healthcare Engineering
Volume 2022, Article ID 7541583, 12 pages
https://doi.org/10.1155/2022/7541583

Research Article
Deep Learning Application for Effective Classification of Different
Types of Psoriasis

Syeda Fatima Aijaz , Saad Jawaid Khan , Fahad Azim , Choudhary Sobhan Shakeel ,
and Umer Hassan
Department of Biomedical Engineering, Ziauddin University, Faculty of Engineering, Science,
Technology and Management (ZUFESTM), Karachi, Pakistan

Correspondence should be addressed to Saad Jawaid Khan; [email protected]

Received 10 October 2021; Revised 20 December 2021; Accepted 31 December 2021; Published 15 January 2022

Academic Editor: Khairunnisa Hasikin

Copyright © 2022 Syeda Fatima Aijaz et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Psoriasis is a chronic inflammatory skin disorder mediated by the immune response that affects a large number of people.
According to latest worldwide statistics, 125 million individuals are suffering from psoriasis. Deep learning techniques have
demonstrated success in the prediction of skin diseases and can also lead to the classification of different types of psoriasis. Hence,
we propose a deep learning-based application for effective classification of five types of psoriasis namely, plaque, guttate, inverse,
pustular, and erythrodermic as well as the prediction of normal skin. We used 172 images of normal skin from the BFL NTU
dataset and 301 images of psoriasis from the Dermnet dataset. The input sample images underwent image preprocessing including
data augmentation, enhancement, and segmentation which was followed by color, texture, and shape feature extraction. Two deep
learning algorithms of convolutional neural network (CNN) and long short-term memory (LSTM) were applied with the
classification models being trained with 80% of the images. The reported accuracies of CNN and LSTM are 84.2% and 72.3%,
respectively. A paired sample T-test exhibited significant differences between the accuracies generated by the two deep learning
algorithms with a p < 0.001. The accuracies reported from this study demonstrate potential of this deep learning application to be
applied to other areas of dermatology for better prediction.

1. Introduction skin cell develops over a period of 28 to 30 days;, however, in

case of psoriasis, this cycle is disturbed and accelerated leading
The skin is important for regulating the body’s temperature and to development of skin cells on the skin surface in 7 days [4].
protects against fungal infection, germs, allergies, and viruses These excess skin cells develop dense, itchy, swollen, red spots
[1]. However, many individuals suffer from skin disorders that in psoriasis lesions which ultimately spread to several parts of
stem from various causes. The most common skin disorders the body. The size of these lesions might range from tiny areas
include eczema, alopecia, ringworm, and psoriasis [2]. Itchy to the full body. The most prevalent forms of psoriasis are
scaly patches which are red in color and most commonly occur pustular, guttate, inverse, plaque, and erythrodermic psoriasis
on knees and elbows represent the symptoms of psoriasis. [5]. These tend to exert a substantial detrimental influence on a
Psoriasis is a persistent skin disorder that cannot be passed person’s quality of life and are often compared to a heart
from one person to another and has no treatment [3]. ailment because they induce depression and are thought to
According to the International Federation of Psoriasis Asso- increase the suicide rate by 30% [6]. Dermatologists usually use
ciations (IFPA), psoriasis affects 125 million people globally, or general observation and biopsies for diagnosis of the correct
around 2% to 3% of the global population [3]. It starts when the type of psoriasis. However, the ambiguity surrounding the
autoimmune system of the body begins to attack skin cells, number of tests required for satisfactory diagnosis regarding
disrupting their regular life and development cycle. Normally, a the adequate type of psoriasis represent the limitations of the
2 Journal of Healthcare Engineering

available diagnostic procedures. Hence, there exist immense Other systems have used multiclass classification along
opportunities for researching new methods in relation to with moth flame optimization for skin lesion segmentation. A
classification and diagnosis of the five types of psoriasis, in- system was developed with a fully automated approach with a
cluding pustular, guttate, inverse, plaque, and erythrodermic CNN model. Method employed HAM1000 dataset, which
[7]. included seven different types of cancerous lesions, including
Machine learning and deep learning approaches have basal cell carcinoma, dermatofibroma malignancy, malignant
demonstrated success in the prediction and categorization of melanocytic, benign melanomas, melanocytic lesions, and
a wide range of illnesses. Deep learning involves the use of actinic keratosis, among others. Classification was carried out
several computer techniques and reflects the ability to learn using a CNN model, which was 90.67% accurate [17]. In
and adapt. For diagnostic purposes, machine learning and another study, machine learning and deep learning techniques
deep learning technologies have been used in a variety of for skin lesion classification and diagnosis were reviewed. It was
medical fields. They have shown accuracy in the diagnosis of concluded that although, machine learning techniques such as
brain tumors, alopecia areata, Alzheimer’s illness, breast k-nearest neighbor (KNN), support vector machine (SVM), k-
cancer, blinding diseases, and renal disease [8–13]. means clustering, and Naı̈ve Bayes methods have been used for
Various deep learning approaches have been used in skin lesion classification, deep learning approaches such as
dermatology to predict and classify skin problems with high CNN tend to outperform machine learning methods [18].
accuracy. For categorizing skin images for the identification of Researchers have also used image augmentation techniques
skin lesions, such as malignant melanoma, basal cell carcinoma, for the identification of skin lesions, in particular, melanoma. A
actinic keratosis, squamous cell carcinoma, and psoriasis, skin skin analysis system was proposed employing the synthetic
analysis algorithms have been developed employing Mask minority oversampling technique (SMOTE) and used the deep
RCNN, transfer learning, and CNN frameworks [14–20]. All of CNN-based SqueezeNet model for classifying malignant skin
these methods entail the classification of a single kind of skin melanoma, atypical nevus, and common nevus from a publicly
condition. Furthermore, to the best of our knowledge, none of available dataset, PH2. The results of the study exhibited an
the deep learning algorithms have been used to classify the five accuracy of 92.18% [19]. In another system, a psoriasis as-
kinds of psoriasis: pustular, guttate, inverse, plaque, and sessment system was proposed using algorithms including
erythrodermic psoriasis. KNN, random forest (RF), deep neural network (DNN), Naı̈ve
Hence, in this paper, we propose a deep learning technique Bayes, and SVM. A total of 80 psoriasis patch images were used,
for the classification of different types of psoriasis as previous and the results demonstrated the highest accuracies of 98.6%
methods have not carried out classification of different types of and 92.6% achieved via RF and KNN, respectively [20]. In a
psoriasis and have used a single deep learning technique. similar study, Dash et al. [21] proposed a CNN model for the
Moreover, previous state-of-the-art works have used different detection of psoriasis. A total of 5241 images of psoriasis lesions
datasets as compared to the datasets that we have used, and we were used, and the findings of the study exhibited an accuracy
have achieved higher accuracies. Our research motivation and of 94.80%. In another study, psoriasis skin image analysis was
proposal exhibit the practical application of deep learning carried out with machine learning methods of KNN, SVM, RF,
approaches for distinguishing and classification of five different and CNN. A total of 90 images of psoriasis skin lesions were
types of psoriasis including pustular, guttate, inverse, plaque, used, and the results demonstrated the highest accuracy of 95%,
and erythrodermic. The results from our study also demon- and 17% being achieved by CNN [22].
strate the future potential of this deep learning application to be The literature review summarized in Table 1 shows that no
applied to further skin disorders and make dermatological work has been done for the identification of the five different
diagnosis more accurate. types of psoriasis, including pustular, guttate, inverse, plaque,
and erythrodermic. Previously, machine learning and deep
2. Related Works learning algorithms have been used to analyze skin images from
publicly accessible datasets. However, no work has been done
Researches have used deep learning approaches for predicting using the Dermnet and Nanyang Technological University
and classifying skin lesions including melanoma and psoriasis. (NTU) databases that we employed in our proposed research to
A smart home system was proposed coupled with sensors and classify five kinds of psoriasis. Furthermore, no study on pso-
artificial intelligence for evaluating skin disorders. The system riasis categorization using both CNN and LSTM deep learning
used normal and melanocytic skin lesion images and applied approaches has been done. As a result, using CNN and LSTM as
CNN to achieve an accuracy of 82.4% [15]. In another study, deep learning techniques, we present an innovative application
extraction and identification of skin melanoma from dermo- for classifying five distinct forms of psoriasis, including pustular,
scopy images was proposed with the help of VGG-SegNet guttate, inverse, plaque, and erythrodermic.
scheme. The technique used CNN, and the results exhibited an
accuracy of 97.16% [16]. In a similar study, skin lesion, in
particular, malignant melanoma recognition method was in-
3. Materials and Methods
troduced using mask region-based convolutional neural net- 3.1. Image-Based Datasets
work (RCNN) and transfer learning-based approach. Three
datasets including ISBI2016, ISB12017, and HAM1000 were 3.1.1. Normal Skin. A total of 172 images of different areas of
used for validation and presented accuracies of 96.3%, 94.8%, body including hands, feet, back, chest, and legs have been
and 88.5%, respectively [14]. collected from the NTU dataset. In particular, the Biometrics
Journal of Healthcare Engineering 3

Table 1: An overview of previously published cutting-edge research.

Year Method Skin disorder Accuracy
2021 CNN [16] Melanoma 97.16%
96.3% (ISBI2016), 94.8% (ISBI2017),
2021 RCNN [14] Melanoma
88.5% (HAM1000)
Dermatofibroma, benign keratosis, basal cell carcinoma, melanoma,
2021 CNN [17] 90.67%
melanocytic nevi, actinic keratosis, and vascular
2021 CNN [19] Melanoma 92.18%
RF and KNN
2021 Psoriasis 98.6% (RF) and 92.6% (KNN)
[20]
2019 CNN [21] Psoriasis 94.80%
2018 CNN [15] Melanoma 82.4%
2018 CNN [22] Psoriasis 95.17%

and Forensics Lab (BFL) NTU dataset was used. The BFL (5) Erythrodermic Psoriasis. A total of 33 images of eryth-
NTU dataset consists of skin images from different parts of rodermic psoriasis were retrieved from the Dermnet dataset
the human body including hands, chest, back, inner forearm, and used in this study. Being one of severe types of psoriasis,
inner thigh, and lower leg [23]. The BFL NTU dataset is a erythrodermic psoriasis involves inflammation with peeling
publicly available dataset with a normalization procedure rashes that burn considerably [29]. Figure 1 represents a
applied so that the aspect ratio and size of each image are the sample input image of erythrodermic psoriasis used in this
same [23]. Figure 1 illustrates a sample image of normal skin study.
used from the BFL NTU dataset.

3.2. Proposed Deep Learning Technique with CNN and LSTM.

3.1.2. Psoriasis. A total of 301 images pertaining to five types The Pandas Python Library used in this study comprises of
of psoriasis have been obtained from the Dermnet dataset. the Dataframe function that aids in organizing the sample
The Dermnet dataset consists of 23 types of dermatological input images and eliminates unwanted rows and columns.
disorders, including plaque, guttate, inverse, pustular, and The code has been written with Python using a Linux
erythrodermic psoriasis. Other types of skin diseases com- workstation involving the use of the TensorFlow package.
prise that of alopecia areata, poison ivy, and eczema [24]. Two deep learning approaches, convolutional neural net-
Figure 1 represents few of the images of each of the five types work (CNN) and long short-term memory (LSTM), are used
of psoriasis that we used in this study. to produce the classification methodology. Figure 2 depicts
the flow process of the proposed deep learning approach. It
(1) Plaque psoriasis. A total of 99 images of plaque psoriasis begins with normal skin and the five kinds of psoriasis,
were retrieved from the Dermnet dataset [24]. Plaque including plaque, inverse, guttate, pustular, and eryth-
psoriasis tends to appear on the skin surface in the form of rodermic psoriasis, as input example images. Following that,
thick and red patches [25]. Figure 1 represents a sample an image enhancement procedure is used to remove dis-
input image of plaque psoriasis used in this study. tortion from the sample images. After the images are en-
hanced, they undergo segmentation after which the images
(2) Guttate Psoriasis. A total of 96 gutatte psoriasis images are divided in training and testing samples. Empirical studies
were collected from the Dermnet dataset and used in this have demonstrated that more accurate and robust results can
study. Guttate psoriasis is a form of skin infection that be acquired via 20% to 30% of the data being used for testing
appears on the skin surface in tear-drop shaped red and itchy and 70% to 80% for training [30]. As a result, 80% of the
patches [26]. Figure 1 represents a sample input image of input sample images are used to train the classification
guttate psoriasis used in this study. model, while the remaining 20% are used for validation and
testing. The categorization of an image into guttate psoriasis
(3) Inverse Psoriasis. Inverse psoriasis also referred to as (class 0), inverse psoriasis (class 1), erythrodermic psoriasis
hidden psoriasis is a form of psoriasis that tends to infect the (class 2), normal skin (class 3), plaque psoriasis (class 4), and
skin folds, areas where one skin region rubs against another pustular psoriasis (class 5) represents the outcome of the
skin region [27]. In this study, we used a total of 25 images of proposed deep learning approach.
inverse psoriasis retrieved from the Dermnet dataset. Fig-
ure 1 represents a sample input image of inverse psoriasis
used in this study. 3.3. Preprocessing of Image

(4) Pustular Psoriasis. A total of 48 images of pustular 3.3.1. Cleaning and Preparation of Dataset. Images retrieved
psoriasis were used from the Dermnet dataset. White bumps from the BFL NTU and Dermnet datasets were cleaned by
filled with pus within or around red scaly patches are opening each image in order to identify that if it is clearly
representative of pustular psoriasis [28]. Figure 1 represents exhibiting the particular diseased part or not and then select
a sample input image pustular psoriasis used in this study. the most authenticated images.
4 Journal of Healthcare Engineering

Dataset

Plaque Guttate Inverse Pustular Erythrodermic

Normal Skin
Psoriasis Psoriasis Psoriasis Psoriasis Psoriasis

172 Images 99 Images 96 Images 25 Images 48 Images 33 Images

Figure 1: Images of normal skin and five kinds of psoriasis, including plaque, guttae, inverse, pustular, and erythrodermic, are shown in the
examples.

Start

Images of 6 types of Psoriasis (Plaque, Inverse, Guttate,

Pustular and Erythrodermic) and normal skin

Image Enhancement and segmentation

Training Sample Testing Sample

Pre-trained Model VGG-19 (CNN) LSTM (Long Short Term Memory) Model

Classification of 6 different classes

End

Figure 2: Flow process of the proposed classification technique.

3.3.2. Data Augmentation. In this study, data augmentation brightness, and pixel luminance values [31]. The sklearn.pre-
was executed by rotating each image to 15°, shifting height and processing package, which is part of scikit-image processing,
width and horizontally flipping the image. We restricted our contains a number of image enhancing algorithms. To improve
data augmentation up to some range, in particular 200 to 400 the sample input images in this research, the image en-
images, so that we could get similar number of images in each hancement method of histogram equalization was used. His-
class. Data augmentation was done in order to overcome bi- togram equalization (HE) improves low-contrast sections of an
asness issues; hence, we increased the size of our data and image, resulting in images with increased overall contrast [32].
applied deep learning techniques on it. Figure 3 exhibits the The HE approach is used in this work to transform RGB images
number of images for each class following data augmentation. into equivalent hue-saturation-value (HSV) image format. In
addition, image segmentation was carried out using the edge
detection approach. The edge detection technique involves
3.3.3. Image Enhancement and Image Segmentation. identifying edges within an image and following change in the
Image enhancement is a technique involving improvement of intensity values, and hence, this results in a segmented image
the input images with overall enhancement of contrast, [33]. The resize function is used in this study to resize input
Journal of Healthcare Engineering 5

Normal Skin 276 Images

Plaque
395 Images
Psoriasis

Inverse
223 Images
Psoriasis Total 1838
Images
Dataset
Guttate
383 Images
Psoriasis

Pustular
300 Images
Psoriasis

Erythrodermic
261 Images
Psoriasis

Figure 3: Dataset following data augmentation.

sample images to a 64 × 64 resolution. Moreover, the anti- (FC) layer is formed with 4096 units along with a softmax output
aliasing technique included in Python’s scikit-image processing layer. Moreover, in this study, the VGG-19 pretrained CNN
package is also employed. In particular, the multisample model is incorporated with some trainable layers in order to
antialiasing (MSAA) technique is used being denoted as true so make the model work more accurately and efficiently. There are
that the rough edges in the input images are smoothened. four convolution layers in the classification model, and a max-
pooling process is preceded by each convolution layer. Such four
convolution layers are used to diagnose the five psoriasis types
3.4. Splitting Dataset. Following image preprocessing, the
through the extraction of features from the input sample images.
data are separated into three portions in an 8 : 1 : 1 ratio,
training, validation, and testing. Within this percentage, 80%
of the images are used to train classification models, with the
remaining 10% used for validation and testing. As shown in 3.6. LSTM Model. As compared to the pretrained CNN
Table 2, 1468 training images of all five classes of psoriasis model used, an LSTM model was trained from scratch as
are used while 182 images are used for validation and 188 only time series-based pretrained LSTM models are avail-
images are used for testing the classification models. able, and we required an image classification model in this
study. In model training of LSTM, we have used 3 × 64 input
layers in order to train the model. The LSTM model com-
3.5. VGG-19 Pretrained CNN Model. A large network visual prises of internal systems called gates that control the in-
geometry group (VGG-19) pretrained CNN model consisting formation flow. Throughout training of the classification
of 19 neural layers is used in this study. The VGG-19 is a deep models, the gates can interpret what data are important [35].
CNN model used to classify images, and the arrangement of the These gates involve sigmoid triggers and tanh activation
layers are shown in Figure 4 [34]. The purpose of max-pooling function as well. Both of these were used in this study, with
is to down sample the expression of the inputs in order to sigmoid squishing around 0 and 1 values, whereas tanh
minimize their computational sizes. Hence, max-pooling is squishes values between −1 and 1 [35]. Sigmoid is used for
responsible for decreasing the volume size. A pretrained CNN output layer calculation, while tanh is used for hidden layers
model was used in this study as for image classification, because of marginally smaller scale as compared to sigmoid.
pretrained CNN models are available, and there was no need Its derivative range is also slightly bigger than sigmoid which
for training the model from scratch. is ideal for steady gradients [35]. Other gates in LSTM we
have used are as follows:
3.5.1. Architecture of VGG-19 Pretrained CNN Model. (i) Forget gate to decide which information should be
Figure 4 exhibits the architecture of VGG-19 pretrained CNN thrown away
model. The model represents that training is done layer by layer
(ii) Input gate to update the state of each cell
in which, convolutional layers with different filter sizes are
present with some pooling layers which are responsible for (iii) Cell state is multiplied by the forget vector point-
reducing the volume for each next layer. Following the com- wise
bination of pooling and convolution layers, a fully connected (iv) Output gate determines the next hidden state
6 Journal of Healthcare Engineering

Table 2: Number of training, testing, and validation images.

Serial number Classes Training images Validation images Testing images
1. Normal skin 220 27 29
2. Plaque psoriasis 316 39 40
3. Inverse psoriasis 178 22 23
4. Guttate psoriasis 306 38 39
5. Pustular psoriasis 240 30 30
6. Erythrodermic psoriasis 208 26 27
Total 1468 182 188

3 3 3 3 3 3 3 3 3
3 3
x x x x x x x x x
x x 3x
M 3 M 3 3 M 3 3 3 M 3 3 3 M F F F S
3 3 3C
I A C A C C A C C C A C C C A C C C O
C C N
N X N X N N X N N N X N N N X F
N N N
P P N P N N P N N N P N N N P 4 4 1 T
N N
U O O O O O 0 0 0 M
T O O O O O 9 9 0 A
12
L 1 L 2 2 L 5 5 5 L 5 5 5 L 6 6 0 X
6 6 8
2 5 5 1 1 1 1 1 1
4 4
8 6 6 2 2 2 2 2 2

Input of 224x224 RGB Image Fully Connected Layer with 1000 and 4096 Units
Convolutional Layers with 3x3 Filters.64,128,152 filters Softmax (Output Layer)
Pooling Layer
Figure 4: Architecture of the VGG-19 pretrained CNN model.

3.7. Extraction of Features. Each input sample image is used build the local representation that finally extracts the texture
to extract attributes, such as color, texture, and shape for our feature. The threshold values are multiplied by the weights
research. The cv2 and skimage Python libraries were used to assigned to the relevant pixels, and the total is calculated to
extract color, texture, and shape data in this investigation. generate an LBP code for the neighborhood pixels. Figure 5
demonstrates how the contrast measure is generated from
the input images. Averaging the grey levels below and above
3.7.1. Color Feature. This study uses the NumPy array
the central pixel is done by subtracting the average values.
function to turn the images into a list of RGB color pixel
Two-dimensional LBP and contrast distributions are most
values. In order to get the average of the three colors (red,
suitable for the texture feature extraction of images.
green, and blue), the cv2 package is used. Initially, the blue
color channel’s mean value is determined, followed by the
mean values for the green and red color channels. As part of
Python’s cv2 module, the NumPy array is able to hold RGB 3.7.3. Shape Feature. These images were used to extract
images in reverse order; hence, each value corresponds to a shape information using the Hu moment shape descriptor.
different color channel. A Python package called OpenCV has been used to import
the Hu moment shape descriptor. The Hu moment shape
descriptor is represented by equations (1) and (2), where S
3.7.2. Texture Feature. Python’s skimage and cv2 libraries denotes the calculated Hu moment and δ represents the
have been loaded into this research in order to take use of the normalized central moment. The central moment is taken
image processing capabilities of scikit. Using local binary into account while calculating Hu moments because it assists
patterns (LBPs), texture descriptors have been used to in the movement of the image’s center region towards the
compute the local representation of a texture feature. The centroid area. Images may be analyzed using Hu moments,
LBP operator integrates statistical and structural models of which measure the contour of the sample input image, and
texture analysis, which have been traditionally considered this results in a NumPy array of the images. In addition, the
distinct [36]. Microprimitives and associated statistical flatten function aids in the creation of the form feature
placement criterions have been used to define texture. A vector by flattening the NumPy array.
supplementary measure of local image contrast may be used
in conjunction with the primitives if desired. This contrast S � δo + δ1, (1)
quantifies the strength of the primitives [36]. Each pixel in
2 2
the image is compared to the surrounding pixels in order to S � (δo − δ1) + 4δ . (2)
Journal of Healthcare Engineering 7

Threshold Multiply

5 4 3 1 1 1 1 2 4 1 2 4
4 3 1 1 0 8 16 8 0
2 0 3 0 0 1 32 64 28 0 0 128

Figure 5: Contrast and LBP calculation.

3.8. Mathematical Operations of CNN and LSTM. Table 3 4.2. Model Evaluation of LSTM. Figure 8 demonstrates the
elaborates the mathematical operations of CNN and LSTM. training and validation of the LSTM model, and it can be
concluded that model accuracy at the time of training was
higher at each epoch as compared to validation. The graph
3.9. SPSS Analysis. IBM SPSS Statistics for Windows, also shows a rise in training and validation, as can be seen.
Version 22.0 of the Statistical Package for Social Sciences Up to the 40th epoch, the model accuracy values for training
(SPSS). Both CNN and LSTM accuracies produced by IBM and validation are almost identical, but after that, training
Corp. were tested using the paired sampled T-test at gains an advantage over validation in terms of accuracy. The
Armonk, NY. Both CNN and LSTM had a total of 30 ac- graph ranges are in between 20% and 100% for training and
curacy samples. 20% and 70% for validation.
Figure 9 represents LSTM model loss and epoch. In both
training and validation, it can be observed that model loss is
4. Results and Analysis decreasing which also reflects that model accuracy, and
4.1. Evaluation of CNN Model. Figure 6 demonstrates the model loss have an inverse relationship as expressed by
training and validation of the CNN model, and it can be equation (4).
concluded that model accuracy at the time of training was LSTM model training was stopped at 100 epochs for
higher at each epoch, whereas during validation, it decreases, model accuracy and 35 epochs for model loss. The reason to
and at some intervals, the model accuracy goes quite low. stop training at these values of epochs was the same com-
The values of the model accuracy per epoch ranges in be- putation of model accuracy and model loss values. This
tween 95% and 100% at the time of training, while at the time means that at 100 epochs and beyond, the same model
of validation, the value lies in between 65% and 80%. accuracy value was being generated. Similarly, at 35 epochs
Figure 7 exhibits CNN model loss versus epoch. This and beyond, the same value of model loss was being gen-
model loss exhibits how well the model is doing in each erated. Furthermore, drop-out layer regularization was used
epoch. It can be observed that as the epoch increases, the for avoiding over-fitting and fine tuning the LSTM model.
validation loss increases while the training loss is quite low. Also, in order to measure the loss in the LSTM model, the
By observing both the graphs simultaneously, it can be binary cross-entropy loss function was used.
observed that when the model accuracy was decreasing the 1
model loss was high, whereas when the model accuracy was Model accuray of LSTM ∝ . (4)
Model loss of LSTM
increasing, model loss was decreasing. Hence, this dem-
onstrates that model accuracy and model loss are inversely
related as represented in equation (3).
The training of the CNN model was stopped at 100 4.3. Performance Evaluation. In the evaluation of CNN and
epochs for model accuracy and 35 epochs for model loss. LSTM, confusion matrices have been employed. Confusion
The reason to stop training at these values of epochs was matrix shown in Figure 10 shows the expected results for
the same computation of model accuracy and model loss each of the six groups. A total of six classifications of
values. This means that at 100 epochs and beyond, the psoriasis have been indicated by the numbers 0, 1, 2, 3, 4, and
same model accuracy value was being computed. Simi- 5, which are guttate, inverse, erythrodermic, normal, plaque,
larly, at 35 epochs and beyond, the same value of model and pustular. True positive is when the classifier correctly
loss was being generated. Furthermore, drop-out layer predicts the positive class, whereas true negative (TN) in-
regularization was used for avoiding over-fitting and fine dicates correct prediction of the negative class by the clas-
tuning the CNN model. Also, in order to measure the loss sifier. False positive (FP) is when the classifiers incorrectly
in the CNN model, the binary cross-entropy loss function predict the positive class while false negative (FN) represents
was used. incorrect prediction of the negative class.
All six classes were evaluated on 188 different images that
1
model accuray of CNN ∝ . (3) were created by the CNN algorithm and shown in Figure 11. Of
model loss of CNN the 188 images analyzed, 158 were correctly categorized,
8 Journal of Healthcare Engineering

Table 3: Mathematical operations of CNN and LSTM.

CNN LSTM
The LSTM deep learning algorithm involves working on a loop
CNNs are feed-forward neural networks in which learning is
network that has two hidden states: cell state and hidden state.
achieved pixel by pixel [37]. CNN employs convolution kernel h, a
Furthermore, it involves assigning weights W as learning parameters
matrix that moves over the input images and executes a dot product
for the classification algorithm [38]. X2 denotes the LSTM layers with
with the central region of the input data, represented by f ∗ h.
the hidden layers being responsible for carrying feedback. The
Following this, the output is yielded as matrix of the dot products
following expression exhibits how the algorithm works with sigmoid
with m columns and n rows and is represented by the following
function σ and the inclusion of past values;
expression;
G (m, n) � f ∗ h (m, n) � 􏽐 􏽐 h[j, k]f(m − j, n − k). Initial Gate � σ (W1Input ∗ X2 + W1Past ∗ Hidden layer),
j k
Input Gate � σ (W2Input ∗ X2 + W2Past ∗ Hidden layer).
Output Gate � σ (W3Input ∗ X2 + W3Past ∗ Hidden layer).

0.14 1.0

0.12 0.9
0.8
Model Accuracy

Model Accuracy
0.10
0.7
0.08
0.6
0.06 0.5
0.04 0.4
0.02 0.3
0.2
0 20 40 60 80 100 0 20 40 60 80 100
Epoch Epoch

train train
validation validation
Figure 6: CNN epoch versus model accuracy. Figure 8: LSTM epoch versus model accuracy.

1.00
0.95 2.5
0.90 2.0
Model Loss
Model loss

0.85
1.5
0.80
1.0
0.75
0.70 0.5

0.65 0.0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Epoch Epoch

train train
validation validation

Figure 7: CNN epoch versus model loss. Figure 9: LSTM epochs versus model loss.

according to the study. This yielded an accuracy of 84.04%. Table 4 shows CNN and LSTM’s performance and
Thirty-four images were classified as guttate psoriasis, 23 as classification results.
inverse, 23 as erythrodermic, 28 as normal skin, 28 as plaque, The graphical illustration of accuracy outcomes of both
and 22 images were classified as pustular psoriasis. the CNN and LSTM models are exhibited in Figure 13.
Figure 12 illustrates the LSTM-created confusion matrix, The results generated by SPSS analysis are demonstrated
which reveals that 136 of the 188 images were correctly by equation (5) where 29 denotes the degrees of freedom,
identified. This yielded an accuracy of 72.34%. Thirty images 20.216 is the t statistic value, and probability value (p value)
were classified as guttate psoriasis, 20 as inverse psoriasis, 19 is less than 0.001, indicating that there is a significant dif-
as erythrodermic, and 26 images as normal skin, 25 as ference between the accuracies obtained via CNN and
plaque, and 16 images were classified as pustular psoriasis. LSTM.
Journal of Healthcare Engineering 9

0 TN FN FP FN FP FN Table 4: Classification results of CNN and LSTM.

1 FP TP FP FN FP FN CNN LSTM
Classes Test Truly Test
2 FP FN TN FN FP FN Truly classified
images classified images
Actual

3 FP FN FP TP FP FN Normal skin 29 28 29 26
Plaque 40 28 40 25
4 FP FN FP FN TN FN
Inverse 23 23 23 20
5 FP FN FP FN FP TP Pustular 30 22 30 16
Erythrodermic 27 23 27 19
0 1 2 3 4 5 Guttate 39 34 39 30
Predicted
Total 188 158 188 136

Figure 10: Confusion matrix.

Accuracy
100.00
0 34 1 1 0 1 2 80.00

1 0 23 0 0 0 0 60.00
(%)
2 1 0 23 0 1 2 40.00
Actual

20.00
3 1 0 0 28 0 0
0.00
4 9 0 0 0 28 3 CNN LSTM

Accuracy
5 4 0 3 0 1 22
Figure 13: Accuracies of CNN and LSTM.
0 1 2 3 4 5

Predicted classification metric that permits evaluation of a model’s

ability to classify true positives of each available class. In
Figure 11: All six classes of CNN images are shown in a confusion order to measure sensitivity, it is necessary to divide the total
matrix. of true positives and false negatives by the number of true
positives [40]. An algorithm’s actual negative rate, or
specificity, helps to identify all of the negative classes that
0 30 0 1 0 5 3 were correctly categorized by the algorithm, as well. The
ratio of true negatives to the total of true negatives and false
1 1 20 0 0 0 2 positives may be described as specificity [40]. It is possible to
determine the accuracy metric by dividing the number of
2 3 1 19 0 1 2 correctly categorized predictions by the total number of
predictions [41]. In this study, the accuracy, sensitivity, and
Actual

3 0 1 0 26 1 1
specificity metrics indicated by equations (6)–(8) were used
4 9 0 2 0 25 4 to assess the performance of the CNN and LSTM models.
Table 5 shows the values of these performance indicators,
5 6 2 2 0 4 16 determined according to their formulae.
TP + TN
0 1 2 3 4 5 accuracy � ∗ 100%, (6)
TP + FP + TN + FN
Predicted
TP
sensitivity � ∗ 100%, (7)
Figure 12: All six classes of LSTM images are shown in a confusion TP + FN
matrix.
TN
specificity � ∗ 100%. (8)
TN + FP
t(29) � 20.216, p < 0.001. (5)

5. Discussion
4.4. Performance Metrics. TP, FP, TN, and FN are computed
with the help of the sklearn library of Python [39]. Classi- 5.1. Main Findings. The goal of this work was to extract
fication methods involve classification metrics namely, characteristics such as color, texture, and form from data-
sensitivity, specificity, and accuracy, that aid in assessing the bases of dermoscopic images of plaque, guttate, inverse,
performance of deep learning algorithms. Sensitivity is the pustular, and erythrodermic psoriasis. The images were
10 Journal of Healthcare Engineering

Table 5: Values of performance metrics including sensitivity, erythrodermic are the five forms of psoriasis that may occur.
specificity, and accuracy for all six classes. Following the extraction of color, texture, and form char-
Algorithm Class Sensitivity (%) Specificity (%) Accuracy (%) acteristics, the convolutional neural network (CNN) and
CNN 0 82.0 71.0 84.2
long short-term memory (LSTM) were used. The application
CNN 1 100.0 95.0 84.2 of CNN presented an accuracy of 84.2% and that of LSTM
CNN 2 85.0 85.0 84.2 presented an accuracy of 72.3%. The accuracies achieved
CNN 3 96.0 100.0 84.2 demonstrate that the proposed deep learning application is
CNN 4 70.0 90.0 84.2 reliable and effective. There are implications for further
CNN 5 73.0 75.0 84.2 research in relation to the existing proposed deep learning
LSTM 0 74.0 65.0 72.3 application which can lead to enhancement of methods in
LSTM 1 86.0 83.0 72.3 biomedical imaging. The existing application can also be
LSTM 2 70.0 79.0 72.3 applied to other skin disorders along with being integrated
LSTM 3 89.0 100.0 72.3 with other deep learning techniques like RNN. Moreover,
LSTM 4 62.0 69.0 72.3
research pertaining to Psoriasis Area and Severity Index
LSTM 5 53.0 55.0 72.3
(PASI) scoring can also be carried out in the future.

classified using deep learning methods such as convolutional Data Availability

neural networks (CNNs) and long short-term memories
(LSTMs). Data is available on request from the corresponding author.
CNN works on pixel-by-pixel convolution, whereas
LSTM runs on feedback mechanism. CNN learns each Conflicts of Interest
minute detail about each pixel value. On the contrary, LSTM
is dependent on past inputs. The achieved accuracy of the The authors have no conflicts of interest in this study.
CNN model as exhibited by Figure 13 is 84.2%, whereas
LSTM accuracy is 72.3%. Figures 6 and 7 corresponding to References
the CNN model accuracy and model loss show that during [1] M. H. A. Fadzil, D. Ihtatho, A. M. Affandi, and S. H. Hussein,
training, CNN model accuracy was increasing and loss was “Area assessment of psoriasis lesions for PASI scoring,”
near to 0%. Whereas in Figures 8 and 9, the LSTM model Journal of Medical Engineering and Technology, vol. 33, no. 6,
accuracy during training increases, while the loss decreases pp. 426–436, 2009.
from 30% to 2%. CNN outperforms LSTM when it comes to [2] S. Kasolang, W. A. Adlina, N. A. Rahman, and N. R. Nik,
categorizing skin images into the five categories of psoriasis “Common skin disorders: a review,” Jurnal Tribologi, vol. 25,
and normal skin, according to our deep learning application. pp. 59–82, 2020.
Other techniques that have used similar deep learning [3] Y. George, M. Aldeen, and R. Garnavi, “A pixel-based skin
methods have also achieved high accuracies. A skin lesion segmentation in psoriasis images using committee of machine
classification system was proposed using plaque psoriasis learning classifiers,” in Proceedings of the 2017 International
Conference On Digital Image Computing: Techniques and
images with classification being executed by CNN, and the
Applications (DICTA), pp. 1–8, Sydney, Australia, December
results of the study reported an accuracy of 60% [22]. 2017.
Psoriasis was classified using dermoscopic images in another [4] A. Rendon and K. Schäkel, “Psoriasis pathogenesis and
investigation. The study’s findings showed that CNN had a treatment,” International Journal of Molecular Sciences,
precision rate of 92.9% [42]. Mathematical processing is the vol. 20, no. 6, p. 1475, 2019.
key to CNN’s improved performance. Accordingly, the [5] Y. George, M. Aldeen, and R. Garnavi, “Automatic psoriasis
84.2% accuracy achieved in this research using CNN is due lesion segmentation in two-dimensional skin images using
to the usage of kernel convolution approach, which converts multiscale superpixel clustering,” Journal of Medical Imaging,
the data into higher dimensions and computes pixel-by- vol. 4, no. 4, p. 1, 2017.
pixel transformation [43]. [6] D. Ihtatho, M. H. A. Fadzil, A. M. Affandi, and S. H. Hussein,
“Area assessment of psoriasis lesion for PASI scoring,” in
The specialty of our proposed deep learning application
Proceedings of the 2007 Tewnty Nineth Annual International
lies in being the first of its kind study encapsulating the Conference of the IEEE Engineering in Medicine and Biology
classification of five different types of psoriasis and normal Society, pp. 3446–3449, Lyon, France, August 2007.
skin. A limitation of this study includes using publicly [7] W. B. Kim, D. Jerome, and J. Yeung, “Diagnosis and man-
available datasets with no collection of clinical data. Hence, agement of psoriasis,” Canadian Family Physician, vol. 63,
future work can be conducted using collected images from no. 4, pp. 278–285, 2017.
clinics and with other deep learning methods so that more [8] L. Sun, S. Zhang, H. Chen, and L. Luo, “Brain tumor seg-
robust classification performances can be accomplished. mentation and survival prediction using multimodal MRI
scans with deep learning,” Frontiers in Neuroscience, vol. 13,
p. 810, 2019.
6. Conclusion [9] C. S. Shakeel, S. J. Khan, B. Chaudhry, S. F. Aijaz, and
U. Hassan, “Classification framework for healthy hairs and
In order to classify the five forms of psoriasis and normal alopecia areata: a machine learning (ml) approach,” Com-
skin, this research used a deep learning classification ap- putational and mathematical methods in medicine, vol. 2021,
proach. Plaque, guttate, inverted, pustular, and Article ID 1102083, 10 pages, 2021.
Journal of Healthcare Engineering 11

[10] G. Lee, K. Nho, B. Kang, K.-A. Sohn, and D. Kim, “Predicting Intelligent Computing and Control Systems, (ICICCS),
Alzheimer’s disease progression using multi-modal deep pp. 1205–1209, Madurai, India, May 2020.
learning approach,” Scientific Reports, vol. 9, no. 1, pp. 1–12, [25] A. W. Armstrong, M. P. Siegel, J. Bagel et al., “From the
2019. medical board of the national psoriasis foundation: treatment
[11] D. Sui, W. Liu, J. Chen et al., “A pyramid architecture-based targets for plaque psoriasis,” Journal of the American Academy
deep learning framework for breast cancer detection,” BioMed of Dermatology, vol. 76, no. 2, pp. 290–298, 2017.
Research International, vol. 2021, Article ID 2567202, [26] E. Errichetti, F. Lacarrubba, G. Micali, A. Piccirillo, and
10 pages, 2021. G. Stinco, “Differentiation of pityriasis lichenoides chronica
[12] D. S. W. Ting, C. Y. Cheung, G. Lim et al., “Development and from guttate psoriasis by dermoscopy,” Clinical and Experi-
validation of a deep learning system for diabetic retinopathy mental Dermatology, vol. 40, no. 7, pp. 804–806, 2015.
and related eye diseases using retinal images from multiethnic [27] A. Zampetti and S. Tiberi, “Inverse psoriasis,” Clinical
populations with diabetes,” JAMA, vol. 318, no. 22, Medicine, vol. 15, no. 3, p. 311, 2015.
pp. 2211–2223, 2017. [28] M. J. Gooderham, A. S. Van Voorhees, and M. G. Lebwohl,
[13] F. Ma, T. Sun, L. Liu, and H. Jing, “Detection and diagnosis of “An update on generalized pustular psoriasis,” Expert
chronic kidney disease using deep learning-based heteroge- Review of Clinical Immunology, vol. 15, no. 9, pp. 907–919,
neous modified artificial neural network,” Future Generation 2019.
Computer Systems, vol. 111, pp. 17–26, 2020. [29] R. K. Singh, K. M. Lee, D. Ucmak et al., “Erythrodermic
[14] M. A. Khan, T. Akram, Y.-D. Zhang, and M. Sharif, “Attri- psoriasis: pathophysiology and current treatment perspec-
butes based skin lesion detection and recognition: a mask tives,” Psoriasis, vol. 6, p. 93, 2016.
RCNN and transfer learning-based deep learning frame- [30] A. Gholamy, V. Kreinovich, and O. Kosheleva, “Why 70/30 or
work,” Pattern Recognition Letters, vol. 143, pp. 58–66, 2021. 80/20 relation between training and testing sets: a pedagogical
[15] D. Połap, A. Winnicka, K. Serwata, K. Ke˛sik, and M. Woźniak, explanation,” Technical Report: UTEP-CS-18-09, University
“An intelligent system for monitoring skin diseases,” Sensors, of Texas, Austin, TX, USA, 2018.
vol. 18, no. 8, p. 2552, 2018. [31] P. K. Verma, N. P. Singh, and D. Yadav, “Image enhancement:
[16] S. Kadry, D. Taniar, R. Damaševičius, V. Rajinikanth, and a review,” Ambient Communications and Computer Systems,
I. A. Lawal, “Extraction of abnormal skin lesion from der- Springer, Singapore, 2020.
moscopy image using VGG-SegNet,” in Proceedings of the [32] H. Singh, Practical Machine Learning and Image Processing,
2021 Seventh International Conference on Bio Signals, Images, Apress, New York, NY, USA, 2019.
and Instrumentation (ICBSII), pp. 1–5, IEEE, Chennai, India, [33] S. C. Shekar and D. Ravi, “Image enhancement and com-
March 2021. pression using edge detection technique,” International Re-
[17] M. A. Khan, M. Sharif, T. Akram, R. Damaševičius, and search Journal of Engineering and Technology, vol. 4, no. 5,
R. Maskeli� unas, “Skin lesion segmentation and multiclass 2017.
classification using deep learning features and improved moth [34] K. Simonyan and A. Zisserman, “Very deep convolutional
flame optimization,” Diagnostics, vol. 11, no. 5, p. 811, 2021. networks for large-scale image recognition,” 2014, https://
[18] M. A. Kassem, K. M. Hosny, R. Damaševičius, and arxiv.org/abs/1409.1556.
M. M. Eltoukhy, “Machine learning and deep learning [35] F. Landi, L. Baraldi, M. Cornia, and R. Cucchiara, “Working
methods for skin lesion classification and diagnosis: a sys- memory connections for LSTM,” Neural Networks, vol. 144,
tematic review,” Diagnostics, vol. 11, no. 8, p. 1390, 2021. pp. 334–341, 2021.
[19] O. O. Abayomi-Alli, R. Damasevicius, S. Misra, [36] Z. Pan, Z. Li, H. Fan, and X. Wu, “Feature based local binary
R. Maskeliunas, and A. Abayomi-Alli, “Malignant skin pattern for rotation invariant texture classification,” Expert
melanoma detection using image augmentation by over- Systems with Applications, vol. 88, pp. 238–248, 2017.
sampling in nonlinear lower-dimensional embedding mani- [37] J. A. Actor, D. T. Fuentes, and B. Rivière, “Identification of
fold,” Turkish Journal of Electrical Engineering and Computer kernels in a convolutional neural network: connections between
Sciences, vol. 29, no. SI-1, pp. 2600–2614, 2021. level set equation and deep learning for image segmentation,”
[20] C.-I. Moon, J. Lee, H. Yoo, Y. Baek, and O. Lee, “Optimization Proceedings of SPIE-The International Society for Optical Engi-
of psoriasis assessment system based on patch images,” Sci- neering, vol. 11313, Article ID 1131317, 2020.
entific Reports, vol. 11, no. 1, pp. 1–13, 2021. [38] S. Liu, C. Zhang, and J. Ma, Edited by D. Liu, S. Xie, and Y. Li,
[21] M. Dash, N. D. Londhe, S. Ghosh, A. Semwal, and Eds., “CNN-LSTM neural network model for quantitative
R. S. Sonawane, “PsLSNet: automated psoriasis skin lesion strategy analysis in stock markets,” in Neural Information
segmentation using modified U-Net-based fully convolu- Processing, D. Zhao and E.-S. El-Alfy, Eds., Springer Inter-
tional network,” Biomedical Signal Processing and Control, national Publishing, New York, NY, USA, 2017.
vol. 52, pp. 226–237, 2019. [39] R. Garreta and G. Moncecchi, Learning Scikit-Learn: Machine
[22] A. Pal, U. Garain, A. Chandra, R. Chatterjee, and S. Senapati, Learning in python, Packt Publishing Ltd., Birmingham, UK,
“Psoriasis skin biopsy image segmentation using deep con- 2013.
volutional neural network,” Computer Methods and Programs [40] J. Brownlee, Machine Learning Mastery with Python: Un-
in Biomedicine, vol. 159, pp. 59–69, 2018. derstand Your Data, Create Accurate Models, and Work
[23] X. Li and A. W. K. Kong, “A multi-model restoration algorithm Projects End-To-End, Machine Learning Mastery, Melbourne,
for recovering blood vessels in skin images,” Image and Vision Australia, 2016.
Computing, vol. 61, pp. 40–53, 2017. [41] J. Hao and T. K. Ho, “Machine learning made easy: a review of
[24] T. A. Rimi, N. Sultana, and M. F. A. Foysal, “Derm-NN: skin scikit-learn package in python programming language,”
diseases detection using convolutional neural network,” in Journal of Educational and Behavioral Statistics, vol. 44, no. 3,
Proceedings of the 2020 Fourth International Conference on pp. 348–361, 2019.
12 Journal of Healthcare Engineering

[42] Y. Yang, J. Wang, F. Xie et al., “A convolutional neural

network trained with dermoscopic images of psoriasis per-
formed on par with 230 dermatologists,” Computers in Biology
and Medicine, vol. 139, Article ID 104924, 2021.
[43] X. Hou, Y. Gong, B. Liu et al., “Learning based image
transformation using convolutional neural networks,” IEEE
Access, vol. 6, Article ID 49779, 2018.