Healthcare Analytics 3 (2023) 100143

Contents lists available at ScienceDirect

Healthcare Analytics
journal homepage: www.elsevier.com/locate/health

An in-depth analysis of Convolutional Neural Network architectures

with transfer learning for skin disease diagnosis
b,∗
Rifat Sadik a, Anup Majumder a, Al Amin Biswas , Bulbul Ahammad a, Md. Mahfujur Rahman b

a
Department of Computer Science and Engineering, Jahangirnagar University, Savar, Dhaka, Bangladesh
b
Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh

ARTICLE INFO
ABSTRACT
Keywords:
Convolutional Neural Network Low contrasts and visual similarity between different skin conditions make skin disease
(CNN) Skin disease recognition a chal- lenging task. Current techniques to detect and diagnose skin disease
Deep accurately require high-level professional expertise. Artificial intelligence paves the way for
learning developing computer vision-based applications in medical imaging, like recognizing dermatological
Xception conditions. This research proposed an efficient solution for skin disease recognition by
MobileNet
implementing Convolutional Neural Network (CNN) architectures. Computer vision- based
Transfer Learning (TL)
applications using CNN architectures, MobileNet and Xception, are used to construct an
expert system that can accurately and efficiently recognize different classes of skin diseases
accurately and efficiently. The proposed CNN architectures used a transfer learning method in
which models are pre-trained on the Imagenet dataset to discover more features. We also
evaluated the performance of our proposed approach with some of the most popular CNN
architectures: ResNet50, InceptionV3, Inception-ResNet, and DenseNet, thus establishing a
comparison to set up a benchmark that will ratify the essence of transfer learning and
augmentation. This study uses data from two separate data sources to collect five different types
of skin disorders. Different performance evaluation indicators, including accuracy, precision,
recall, and F1-score, are calculated to verify the success of our technique. The experimental
results revealed the effectiveness of our proposed approach, where MobileNet achieved a
classification accuracy of 96.00%, and the Xception model reached 97.00% classification accuracy
with transfer learning and augmentation. Moreover, we proposed and implemented a web-
based architecture for the real-time recognition of diseases.

1. Introduction
Most skin diseases have revealing symptoms such as rash,
1.1. Background ulcers, lesions, moles, etc. However, the diagnosis of skin
diseases faces some difficulties. The most common obstacle
Skin is the most vital and sensitive organ in the human is that many skin conditions have similarities between them
body, shielding against heat, injury, and infections. that are not distinguishable visually. Besides, symptoms are
Unfortunately, the skin condition is sometimes disrupted due constantly changing over a long process. Even physicians are
to bacterial and viral infection, fungus, lack of a strong bound to visual imperfections due to the lighting con- ditions
immune system, and genetic imbalances. In many cases, of the environment, the skin color of the patient, and their
diseases caused by those factors have macabre effects on professional experience. In most cases, early detection of
human life. In addition, some skin diseases are contagious, skin diseases reduces the risk factors. The mortality rate of
risking not only individuals but also others related to the some diseases with a high mortality rate can be reduced to
infected. Statistics [1] reported that over 100 million people 90% if diagnosed in the early stage [3].
all over the world are suffering from different types of skin
indispositions; the most frequent skin disor- ders are Atopic
dermatitis, Eczema, Herpes, Nevus, Warts, Ringworm, 1.2. Motivation
Chickenpox, and Melanoma, etc. American Cancer Society
reported [2] that, by the end of the year 2020, 100,350 Researchers are actively investigating methods to develop
new melanoma cases will be reported and diagnosed, and skin dis- ease recognition systems. Many studies have
almost 6850 people are about to die because of melanoma. utilized image process- ing techniques incorporating
∗ Corresponding author. statistical analysis to extract information

E-mail addresses: rifat.sadik.rs@gmail.com (R. Sadik), anupmajumder@juniv.edu (A. Majumder), alaminbiswas.cse@gmail.com

(A.A. Biswas), bulbul@juniv.edu (B. Ahammad), mrrajuiit@gmail.com (Md.M. Rahman).

https://doi.org/10.1016/j.health.2023.100143
Received 11 June 2022; Received in revised form 25 November 2022; Accepted 24 January 2023

2772-4425/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

about skin conditions [4–8]. Researchers were trying to Besides, we proposed and implemented a web-based
recognize skin diseases by analyzing textures, structures, architecture for the real-time recognition of diseases. We
and colors in these approaches. Methods like Self-organizing deployed our trained models on the web using Flask
Map (SOM), Radial Basis Function (RBF), Gray Level Co- framework [27], and the recognition of skin diseases can be
occurrence Matrix (GLCM), etc., were used for such done remotely using this system. Our proposed approach can
approaches. But all these methods lack in terms of precision aid health professionals by recognizing different skin diseases
and accuracy since these methods require sufficient data, more efficiently and making the diagnosis process more user-
good coverage of the input space, and high dependence friendly for the patients. Moreover, besides pandemics and
on texture features such as contrast, correlation, entropy, natural disasters, a cloud-based healthcare system can be
etc. built to operate the healthcare system remotely. Here we sum
In recent times, Artificial Intelligence (AI) has evolved up the whole concept of this work’s contribution below:
enormously in the clinical context, or medical field [9,10]. In
the medical field, Machine learning (ML) and Deep Learning
(DL) algorithms prove their worth in implementing smart
and automated AI-based systems [10–13]. Researchers have
pulled their strings to develop more ad- vanced
frameworks that can be applied in various image-based ap-
plications. Convolutional Neural Network (CNN) is
considered the state-of-the-art method in the analysis of
visual imagery. In medical image analysis such as X-ray
images, MRI images, CNN model and its derivations such as
ResNet, VGG-16, GoogleNet, AlexNet, etc., have shown
significant results in detection, recognition, and classification
tasks [14]. However, deep learning architectures like CNN
require immense computation resources as well as a lot of
image data to train the proposed model [15]. Due to the lack
of sufficient data and resources, the field of medical image
analysis for skin diseases is yet to explore to the full extent.
Pretrained CNN models have come to a point by researchers
to aid the purpose. Besides, image analysis techniques such
as Augmentation are widely used to construct a generalized
model and robust systems where training data is
inadequate.
CNN architectures like MobileNet and Xception are
helping re- searchers to bring out new intelligent systems
nowadays. For exam- ple, the MobileNet model shows high
accuracy for the classification task [16] where welding
defects from images were analyzed. In medical imaging, such
as children’s colonoscopy [17] combination of MobileNet
with DenseNet is proposed for better classification results. In
[18], lung diseases were analyzed and detected from chest
X-ray images using the MobileNet model. In language
processing tasks [19,20] MobileNet model was studied for
the recognition task of Bangla characters which are
handwritten and complex sign language translations. The
Xception model is also widely used for different computer
vision-based tasks. For example, chest X-ray images were
analyzed using the Xception model in [21,22] to differentiate
between COVID-19 lung condition and normal pneumonia. In
[23], Xception based framework is used to clas- sify and
authenticate forensic images. Researchers also implemented
this model for the garbage image classification task in [24]
for the productive garbage management system.

1.3. Contribution

In this work, we implement an automated system based

on com- puter vision-based techniques where two structured
Convolutional Neu- ral Network architectures MobileNet
[25] and Xception [26], con- tribute to the recognition of
different types of dermatological diseases, namely Atopic
dermatitis, Eczema, Herpes, Nevus, and Melanoma. In order
to construct an accurate model, we combined these two
archi- tectures with transfer learning and a real-time image
augmentation process. In addition to that, we evaluated the
effectiveness of our propositions by comparing the
performance with state-of-the-art deep learning models such
as ResNet50, InceptionV3, Inception-ResNet, and DenseNet.

2
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

• Propose an automated framework for skin disease median, smooth filter, binary mask, histogram, YCbCr, etc.,
recognition based on pre-trained CNN architectures, were used for feature extraction. An artificial neural net-
namely MobileNet and Xception. work (ANN) was used for training and test purposes. On a
• For a more robust and generalization model, real-time dataset, the proposed model obtained a
augmentation and transfer learning techniques are classification accuracy of 90%. To make a classification
included. between skin conditions such as normal, spots, and wrinkles,
• Propose and implement a web-based application to Jhan S. Alarifi et al. [29] used traditional ML approaches
recognize skin diseases remotely. based on SVM and CNN. SVM used feature extraction
• Evaluate the model’s performance by comparing it with techniques like LPB and HOG. For CNN, GoogleNet
other deep learning models such as ResNet50, architecture was implemented with different optimizers. The
InceptionV3, Inception- ResNet, and DenseNet. experimental result showed that GoogleNet with NAG
optimizer outperformed SVM in all aspects, reached to an
2. Literature review accuracy level of 89%. Yuexiang Li and Linlin Shen [30]
proposed
Researchers were trying to develop an efficient and
effective system that visually recognizes different classes of
skin diseases. Some of the approaches include image
processing techniques with statistical methods, texture, and
color analysis. AD Mengistu, DM Alemayehu [4] proposed
image processing techniques for recognizing and predicting
skin cancers. Predefined classes of skin cancers collected
from the American cancer society and DERMOFIT were
used in this experiment. A hybrid method that integrates
two image processing techniques, namely a Self-organizing
map (SOM) and radial basis function (RBF), was used in this
recognition task, and image features such as color, tex-
tures, and image structure were combined. Further, the
acquired results were compared with other approaches such
as KNN, Naïve Bayes, and ANN. The reported result
revealed that the overall accuracy for this applied hybrid
method was 93.15%. Manish Pawar et al. [5] Identify
different skin disease conditions based on feed-forward
backpropaga- tion neural networks. Texture features were
used as key attributes for image recognition purposes that
were analyzed from the GLCM method. Three skin
conditions were selected for the classification task, and the
overall accuracy was reported at 66.66%. To enhance the
scope for identifying multiple skin diseases, Li-sheng et al.
[6] pro- posed a method that combines both color and
texture features. The preprocessing task included noise and
background removal through filtering and transformations.
The GLCM approach was implemented to extract texture
features such as contrast, correlation, entropy, etc., and for
color feature extraction watershed algorithm was used.
For this research purpose, three types of common skin
diseases, namely herpes, dermatitis, and psoriasis, were
classified using a support vec- tor machine (SVM) classifier.
The average accuracy while recognizing those 3 classes of
skin disease images reached 90% using SVM classifier and
combining color and texture features. Md. Nazrul Islam et
al. [7] established a system for recognizing multiclass skin
diseases that relied on image texture. Different
preprocessing operations, such as resize, grayscale
conversion, contrast enhancement, and noise removal were
conducted for this experiment. Images textures were
extracted using the GLCM method, and segmentation was
carried out using Maximum Entropy Thresholding. Finally,
the Backpropagation (BPN) algorithm was used to classify 3
different classes of skin disease images Eczema, Impetigo,
and Psoriasis. The obtained accuracy for this method was
reported at 80% along with sensitivity and specificity of
71.4% and 87.5%, respectively.
Rahat Yasir et al. [28] proposed a computer vision-based
approach for recognizing skin diseases from images.
Different preprocessing algo- rithms, like sharpening,

3
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

deep learning methods for tasks like segmentation, ISIC-

extraction, and classification, which are involved in skin 19. The proposed experiment has achieved a specificity of
lesion detection. A fully convolutional residual network 91% with the ISIC-17 dataset. Reis et al. [38] presented a
(FCRN) framework with a lesion index calculation unit CNN network based on inception block (InSiNet) to detect
(LICU) was used for segmentation and classification. On the lesions. The proposed method reported an accuracy of 94.59%
other hand, feature extraction was carried out using the with the ISIC 2018 dataset. A details comparison between
Lesion Feature Network (LFN) framework. The experiment different CNN architectures was demonstrated in this study
was evaluated on ISIC 2017 dataset containing 2000 images and InSiNet outperforms all of them. To classify skin cancer,
for training and 150 images for validation. The accuracies of Gupta et al. [39] used the CNN model that can work on
the proposed approaches for segmenta- tion and both dermoscopic and photographic images. This approach
classification were 75% and 91%, respectively, and the obtained a classification accuracy of 83.2% using the
feature extraction task achieved an accuracy of 84%. An Inception V3 model.
automated image recognition system was proposed by Kalaiyarivu and Nalini [40] studied a CNN-based approach
Jainesh Rathod et al. [31] for detecting skin diseases. Noise to detect skin diseases by extracting color and texture
removal and image enhancement filters were used in the features. The proposed
preprocessing phase. CNN algorithm was applied to the
DermNet dataset as a feature extractor and classifier. An
accuracy of 70% was reported in this experiment.
Min Chen et al. [32] proposed a real-time and dynamic
framework for skin disease recognition that is composed of
self-learning with a wide collection of data for effective
interaction between users. A data filter algorithm was
employed for the removal of unwanted data and feeding the
network with valuable data. Information entropy was used
for this filtering task. Three CNN learning models, namely
LeNet-5, AlexNet, and VGG-16, were used for the
classification and prediction task. The authors also test the
reliability and validity of the proposed system by analyzing
the computation and transmission delays of the system. This
analysis, it had been shown that the communication delay of
the AlexNet and LeNet models is 75 ms and 63 ms,
respectively. Md Ashraful Alam Milton [33] experimented
with Melanoma detection techniques where different deep
neural networks like PNASNet-5-Large, InceptionResNetV2,
SENet154, InceptionV4 were used. Images from ISIC 2018
dataset were used to train and test the proposed models. All
the images were preprocessed by several operations such as
normaliza- tions, and augmentation before training.
Parameters were initialized using the ImageNet model, and
models were fine-tuned. The highest validation score was
76% which was reported for the PNASNet-5-Large model.
For the construction of an automated computerized diagnosis
system, Haofu Liao [34] proposed a method based on deep
CNN. In this study, advanced CNN architecture such as
VGG-16, VGG-19, and GoogleNet was implemented. The
experiment was conducted on two different datasets, namely
Dermnet and OLE, and the performance of the models was
compared. All the models that were used in this study were
pretrained on the ImageNet dataset. On the DermNet
dataset, the top-5 accuracy was 91% using the VGG-16
model, while for the OLE dataset, the top-5 accuracy
reached 69.5%. Shanthi et al. [35] suggested a method that
was used to detect four types of skin diseases from the
DermNet Dataset. The CNN architecture called ALexNet,
which is utilized by 11 layers, was used for the detection
task. The maximum pooling layer with a learning rate of
0.01 was used for the model training purpose. The highest
accuracy was 93.3% for the Eczema herpeticum class.
Srinivasu et al. [36] combined MobileNet V2 with LSTM to
classify skin disease. HAM10000 dataset was used in this
experiment and the reported accuracy was 85%. They have
also proposed a web application for the classification of skin
decreases. In another work, Iqbal et al. [37] proposed multi-
class classification for skin diseases using a deep CNN model
and used 3 different datasets, namely ISIC-17, ISIC-18, and

4
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
acquired from a large dataset is applied. The domain
knowledge from this phase is then transferred to the
MobileNet and Xception model in the building phase. An
unlabeled image data is then fed to the learned network.
Finally, the model generates the class labels based on the
knowledge it has gained from the previous phase. The
schematic representation of our approach is given in Fig. 2.
The required steps with the setup that will be carried out
throughout the experiment are illustrated in Table 1.

Fig. 1. Architecture of the expert system to recognize skin

diseases.

CNN model achieved an accuracy of 87.5%. In another

work, Kousis et al. [41] used 11 different CNN models to
recognize skin cancer. In this approach, they have used the
HAM10000 dataset, and they reported an accuracy of
92.25% using the DenseNet169 model. Ahmad el at. [42]
proposed a hybrid classification approach using CNN and
stacked BLSTM. In this work, BLSTM is used for feature
extraction and then ensembled with a deep CNN network
for the classification task. The authors experimented on two
different datasets, one customized and another one
HAM10000, and reported an accuracy of 91.73% and
89.47%, respectively. Aijaz et al. [43] proposed a deep
learning-based application where different categories of
skin diseases are classified. two different deep learning
models CNN and LSTM were used in this approach. For
better results, different pre-processing techniques such as
augmentation, enhancement, and segmentation were
employed in this study and achieved an accuracy of 84.2%
and 72.3% for CNN and LSTM, respectively.

3. System architecture and research methodology

This section scrutinizes the pertinent technologies and

architectures that will be used to develop an automated
system for different skin disease recognition.

3.1. System architecture

We presented a web-based medical expert system using

deep learn- ing framework for recognizing different skin
diseases. The proposed overview of the system is illustrated
in Fig. 1. The hypothesis behind this architecture is that a
user captures an image of the diseased area using a smart
device where the proposed application will be pre- installed.
Then the image will be sent to the expert system
through the application. Then feedback will be generated
based on our trained model or expert system. The feedback
will be returned to the user who seeks to identify or
diagnose skin diseases via email or SMS.

3.2. Research methodology

Since we have minimal images to train our deep learning

model, we propose models with real-time data augmentation
and transfer learning approach integration. First, the
acquisition of pretrained weights from tasks conducted on
the ImageNet dataset is made in the building phase.
Features from the training data are extracted as well as
the labels in this phase. Second, a feature tensor combines
these features according to the class labels. Third, the
transfer learning approach uses pretrained weights that are

5
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 2. Proposed system architecture. (A systematic representation of our proposed approach including data acquisition, preprocessing using augmentation,
transfer learning, training, testing, and predictions carried out in building and deployment phases.)

Table 1
All the necessary steps with the setup that will be carried out throughout the
experiment. Algorithm: Experimental setup

Input 1. Collected images of 5 classes of skin diseases .

Environment 2. Google Colab.
3. Import all necessary libraries and packages .
Configuration
4. Import the images.
Directories Configuration 5. Construct directories for Training, testing and validation.
6. Build CNN models. For transfer learning use a model trained on ImageNet dataset.
Training and Testing
7. Fine-tune the models by adding additional global average pooling layers, a fully connected layer and Softmax class
8. Model compile with RMSProp optimizer and learning rate of 0.001.

Model Compilation 9. Set 100 epochs for model fitting.

10. Use val_accuracy monitor as model checkpoint.
11. Save model.

Performance Evaluation 12. Generate classification report and confusion matrix.

13. Generate AUC–ROC curve.
14. Generate model accuracy and loss reports.

Prediction 15. Load best model.

16. Load random images.
17. Predict disease classes.

We have implemented six different CNN-based differentiating different classes of images. The feature
architectures namely ResNet50, InceptionV3, Inception- detection task is the backbone of the CNN model, which has
ResNet, DenseNet, MobileNet, and Xception. But we focused been carried out using the feature extractor filter or Kernel.
more specifically on MobileNet and Xception Model. The
remaining models are used in this study to compare the
performance of our propositions.

3.2.1. Convolutional Neural Networks (CNN)

CNN is the most popular artificial neural network
specially designed for computer vision-based applications
that incorporates analyzing vi- sual imagery [44]. The
network takes an image as input and processes the image for
extracting different features and patterns from that input
image. These features are also made distinguishable by the
network. Both spatial and temporal characteristics are
captured using CNN. These characteristics are used in

6
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
The learning process of CNN constitutes convolutional
layers, non- linear processing units, and layers for
subsampling tasks [45]. The working of CNN implements a
layered architecture and presented in Fig. 3. Three main
layers, namely convolution, pooling, and fully con- nected
layer, are used to build a CNN model [46]. Convolutional
layers have a convolutional kernel that works as a feature
extractor. These kernels slice the input image into receptive
fields. The relation between the input feature map and
output feature map can be expressed using convolutional
∑ ∑
operation, i.e., 𝐹 (𝑥, 𝑦) = (𝑓 ∗ 𝑘)(𝑥, 𝑦) = 𝑓 (𝑖, 𝑗)𝑘(𝑥 − 𝑖
𝑖, 𝑦 − 𝑗). 𝐹 (𝑥, 𝑦) and 𝑓 (𝑥, 𝑦) corresponds to the output and input
feature map, and k(x,y) represents the element of the
corresponding kernel. The pooling layer involves an
operation that sums up all the relevant and similar
information from the neighborhood. The size of the input
feature map has been reduced by cutting down the number
of param- eters. The pooling operation can be formulated
using the equation,
𝑍 = 𝑔𝑝(𝑓 ) where Z is the polled feature map operating with
input

7
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 3. A convolutional neural network (CNN) architecture with its dimensions. (A layered representation of CNN architecture for performing different
operations like convolution, pooling, and consist of convolution layers, pooling layers and fully connected layer.)

Fig. 4. Architecture of MobileNet. (A CNN architecture performing Depthwise and Pointwise convolution on the input image for the completion of the
filtering task and the creation of linear output combinations.)

feature map f. Finally, the classification task has been 𝑀 of feature map F and produces 𝐷𝐺 × 𝐷𝐺 × 𝑁 of feature map G.
carried out using a global operation carried out in a fully The value of 𝐷𝐹 × 𝐷𝐹 represents the dimension (height*width) of
connected layer (FC). All the extracted features are analyzed the input image and 𝐷𝐺 × 𝐷𝐺 represents the dimension
in this layer and create a non-linearity between them. (height*width) of the output image. Here 𝑁 is the number of
input channels or input depth, and M is the number of output
3.2.2. MobileNet channels or output depth. For standard
MobileNet is a popular Deep CNN network, widely used in
computer vision-based applications such as image
classification, categorization or segmentation, etc., for its
lightweight and small architecture and fast operational
characteristics [25]. The fabrication of MobileNet is
established on depthwise separable filters represented in
Fig. 4. The main focus of this model is to optimize latency
with a small network and make a model that is suitable for
deploying on mobile devices. Mo- bileNet architecture is
incorporated with two steps, namely depthwise convolutions
and pointwise convolutions. First, the feature extraction
process is carried out by depthwise convolutions, where
only a filter processes each input channel. Then the
pointwise 1 × 1 convolution is applied that combines
features obtained from depthwise convolutions. In depthwise
separable convolutions, extraction of features, and com-
bining those features are done by separate layers. This
results in the reduction of computation time and
computation cost, and model size. There exist some
architectural differences between the general con- volutional
layer and the depthwise convolutional layer. The input that
is taken by a standard convolutional layer can be expressed as
𝐷𝐹 ×𝐷𝐹 ×

8
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
convolution layers where K is the kernel of size 𝐷𝐾 ×𝐷𝐾 ×𝑀
×𝑁 where
𝐷𝐾 × 𝐷𝐾 denotes the dimension of the kernel. The output
feature map is given by the following equation
∑
𝐺(𝑘,𝑙,𝑛) = 𝐾(𝑖,𝑗,𝑚,𝑛) .𝐹(𝑘+𝑖−1,𝑖+𝑗−1,𝑚)
(1)
(𝑖,𝑗,𝑚)

For the depthwise convolution layer, is depthwise

convolution ker- nel is denoted by 𝐾̂, and the size of
this kernel can be computed as
𝐷𝐾 × 𝐷𝐾 × 𝑀. So the depthwise convolution for input
depth can be written as
∑
𝐺̂(𝑘,𝑙,𝑚) = 𝐾̂(𝑖,𝑗,𝑚) .𝐹(𝑘+𝑖−1,𝑖+𝑗−1,𝑚) (2)
(𝑖,𝑗)

Here 𝑚𝑡ℎ filter in 𝐾̂ applied to the 𝑚𝑡ℎ channel in F to

produce the 𝑚𝑡ℎ channel of 𝐺̂. The total computational
cost for depthwise convolutions is given by 𝐷𝐾 .𝐷𝐾 .𝑀.𝐷𝐹 .𝐷𝐹

3.2.3. Xception
Xception is another class of Deep CNN which is adapted
from the Inception-V3 model [26]. The model is constructed
based on the intu- ition of the depthwise separable
convolutional module. Modification is made in the inception
block of the Inception-V3 model. The modified architecture
for Xception has a wider inception block than Inception- V3.
It has spatial dimensions of 1 × 1, 5 × 5, and 3 × 3, which is
replaced in the Xception model with a single dimension of
size 3 × 3 and 1 × 1, i.e., Convolution part is divided into
spatial and pointwise convolution. Fig. 5 illustrates the
architecture of the Xception network. Firstly, 1 × 1
pointwise convolution is applied, and then a 3 × 3 depthwise
convolution is applied [45]. This approach results in the

9
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 5. Architecture of Xception. (A layered architecture of Xception consisting of 36 convolutional layers and 14 modules. It implements a 1 × 1
pointwise convolution followed by a 3 × 3 depth-wise convolution.)

Fig. 6. The process of transfer learning. (Pretrained weights from earlier tasks conducted on a very large dataset has been used for the purpose of
transporting knowledge. An additional global average pooling layer, a fully connected layer, and a Softmax layer are added for fine-tuning the network.)

reduction of parameters and layers and makes the network

lightweight. Disengagement of this correlation is followed by a lot of images and computational resources [47]. To
Eqs. (3) and (4). overcome this, deep learning models can utilize the TL
approach, in which a model
∑
𝑓𝑘 (𝑝, 𝑞) = 𝑓 𝑘(𝑥, 𝑦).𝑒𝑘(𝑢, 𝑣) (3) that has been trained for one task can be used as a
baseline model
(𝑙+1) 𝑙 𝑙 for another. This method of reusing models that have been
(𝑥,𝑦)
trained previously with a large amount of data can be used in
(𝑙+1) another training
𝑘
𝐹 (𝑙+2) = 𝑔𝑐 (𝐹 𝑘 , 𝐾(𝑙+1) ) (4) process that has a small amount of data and paves the way for
achieving
Here, F corresponds to the feature map of l integration of features.
transformation layers, (𝑥, 𝑦) and (𝑢, 𝑣) show the spatial indices
of feature map F and kernel K having depth one. Kernel K is 3.3. Transfer learning
spatially convolved across feature map F. Here 𝑔𝑐 (.) indicates
the convoluted operation. In total, a basic Xception model has The Transfer Learning (TL) approach in the context of deep
36 convolutional layers and 14 modules. Among these, 12 learning is a pervasive method in computer vision-related
modules are connected with a residual layer boosting the tasks. However, creat- ing a robust and generalized deep
merging process and paving the way for achieving higher learning model, it is highly required
accuracy. Architecturally, the Xception network consists of 3
flows, namely Entry flow, Middle flow, and Exit flow.
Downsampling of input images with dimensionality reduction
is carried out using the Entry flow part. Learning from
features and optimizing those features is done by the Middle
flow part of the network. Finally, the Exit flow carries out the

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
higher accuracy [48]. In general, weights are initialized
using random numbers in the training process of neural
networks. These assigned weights are then slowly updated
during the training process. So in most cases, training with
a small number of training data cannot achieve sufficient
accuracy. To perform the transfer learning process, we
should prepare a neural network model trained with many
data that can handle similar types of data, which becomes
the source model for transfer learning.
In the transfer learning process, features learned from
huge image sets such as ImageNet are highly transferable
to a variety of image recognition tasks [49]. This process is
depicted in Fig. 6. Several ways to transfer knowledge from
one model to another. One approach is to train the top layer
of the already pretrained model and then replace it with a
randomly initialized one. After that, the top layer
parameters are trained for the new task while all other
parameters remain fixed. This approach best suits a task
where there is a maximum similarity between the pretrained
model and the new task. If we have more data, then we can
train the entire network by unfreezing these transferred pa-
rameters. Only the initial values of the parameters are
transferred while

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 7. Different skin diseases used in our approach.

Table 2 the model, we performed real-time image augmentation in

our study,
Overall dataset splitting.
expanding our training data virtually. The augmentation
Skin diseases Training images Validation images Test images
operations that are performed in this task are flip, shift, and
Atopic dermatitis 2610 432 100
zoom. Both vertical and horizontal flips are performed to
Eczema 4750 950 100
Herpes 4200 840 100
reverse the pixel columns or rows. Shift operation moves all
Nevus 1955 391 100 of the pixels unidirectional, and using zoom operation,
Melanoma 1720 344 100 images are randomly zoomed by adding new pixel values.
Total 15235 2957 500 Augmentation techniques that are used in our approach
are illus- trated in Fig. 8.

4.4. Performance evaluation metrics

weights are initialized using pretrained models instead of
initializing them randomly, boosting the convergence
The performance of a classifier is described through the
process.
confusion matrix, which gives an insight into the correct and
incorrect predictions made by the classifiers [53]. A classifier
4. Experimental evaluation and result analysis
is used to predict some classes that can be either true or
false. There can be four cases as output while classifying
4.1. Environment specifications
some data belonging to more than one class. Firstly all the
predictions (true or false) are correct, which is indicated by
Image analysis or classification requires intense True Positive (TP) and True Negative (TN). However, there
computing powers, and GPU (Graphics Processing Unit) can can be another case in which the prediction is true, but in
provide such computing com- patibility. But GPU installation reality, it is false, and vice- versa. These two cases are called
is expensive and requires additional hardware to support the False Positive (FP) and False Negative (FN). Not only that,
computing task. So we use the Google Colab 1 platform to we can calculate some more specific metrics from the
train our model, which provides us with high-end GPU on the confusion matrix that can be deciding factors for revealing
cloud. It comes with all the necessary packages which are the classification performance of our models. These metrics
used in the training process, so there is no burden of are Accuracy, Precision, Recall, and F1-score. These metrics
installing packages or extra storage [50]. Google Colab are calculated using the following formulas.
comes with NVIDIA K80 GPU, GPU memory of 12 GB, Up to Accuracy: Accuracy is the indicator of how well a model can
2.91 teraflops double-precision rendition, and disk space of predict true and false classes precisely and expressed using
∑
358 GB. These specs give an enormous computation formula (5).
environment to train Deep Learning models. 𝑁
𝑀
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = ∑𝑖 𝑁𝑖 × 100% (5)
𝑇
4.2. Dataset description 𝑖 | 𝑖|
∑𝑁
where,
∑ indicates the total number of correct
We have used 5 classes of skin diseases, namely Atopic
|𝑇𝑖 | 𝑀
predictions, and
dermatitis, Eczema, Herpes, Nevus, and Melanoma. Since 𝑁
is𝑖 the total number of predictions.
there is no available dataset that contains images of all these 𝑖
When it comes to binary classification, Accuracy is
classes, we prepared our dataset by collecting images from represented using the following formula (6)
two different sources. We have collected images for Atopic
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 𝑇𝑃 +𝑇𝑁 (6)
dermatitis, Eczema, Nevus, and Herpes from Dermnet [51]. = 𝑇𝑃 +𝑇𝑁 +𝐹𝑃 +𝐹𝑁
For Melanoma images, we have used the HAM10000 dataset
where, 𝑇 𝑃 = True Positives, 𝑇 𝑁 = True Negatives, 𝐹 𝑃 = False
[52]. A total number of 18692 images are used in our
Posi- tives, and 𝐹 𝑁 = False Negatives.
approach, split for training, validation, and testing purposes. Precision: Precision indicates how well a classier performs
A glimpse of images constituting our dataset is given in Fig. in terms of predicting correct outcomes that are positive.
7 Splitting the dataset into training and testing datasets Mathematically repre- sentation can be established using
depicts in Table 2. the formula (7)
𝑇𝑃
4.3. Data preprocessing 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (7)
𝑇𝑃 +𝐹𝑃
The proposed CNN architecture MobileNet and Xception Recall: Recall indicates the performance of a classier by
require very less preprocessing images as they extract measur- ing the proportion of true positive observations that
features directly from were correctly predicted. Formally Eq. (8) defines Recall,

1
R. Sadik, A. Majumder, A.A. Biswas et 𝑇𝑃 Healthcare Analytics 3 (2023)
images. MobileNet model requires an input shape of 224 × 𝑅𝑒𝑐𝑎𝑙𝑙 = (8)
224, and the Xception model requires images of dimension 𝑇𝑃 +𝐹𝑁
229 × 299. So firstly,
images are resized according to the measurement for each F1 score (F-measure): F1 score is the symphonic average
model. Since a robust model requires many images to train of pre- cision and recall. Formally it is represented
and validate mathematically as Eq. (9)

2 ∗ 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙 (9)

1 𝐹 1 𝑠𝑐𝑜𝑟𝑒 = +
𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑅𝑒𝑐𝑎𝑙𝑙
https://colab.research.google.co
m/

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 8. Augmentation techniques that are used in our approach.

Table 3 Table 4
Class-wise classification results of MobileNet and Xception. (Values of Class-wise classification results of MobileNet and Xception. (Values of
evaluation metrics Precision, Recall, and F1-score for MobileNet and evaluation metrics Precision, Recall, and F1-score for MobileNet and Xception
Xception model with Transfer Learning approaches is presented for each model with without Transfer Learning and without augmentation approaches is
disease classes.) presented for each disease classes.)
Method Class Recall (%) Precision (%) F1 (%)
Model Class Recall (%) Precision (%) F1 (%)
Atopic dermatitis 88.3 91.0 89.6
Atopic dermatitis 97.00 90.70 93.71
Eczema 89.00 95.70 92.22
MobileNe Herpes 95.00 96.94 95.96 Eczema 85.7 66.0 74.5
t
Melanoma 100.00 97.08 98.51 MobileNe Herpes 84.6 99.0 91.2
Nevus 99.00 100.00 99.50 t Melanoma 89.4 93.0 91.1
Nevus 100. 99.0 99.4
Atopic dermatitis 96.00 97.00 96.50
Eczema 90.00 95.74 92.80 Atopic dermatitis 84.2 91.0 87.4
Herpes 99.00 92.52 95.65 Eczema 91.0 71.0 79.7
Xception
Melanoma 100.00 100.00 Herpes 80.4 99.0 88.7
100.00 Xception Melanoma 97.8 93.0 95.3
Nevus 100.00 100.00 Nevus 100.0 96.0 97.9
100.00

Sometimes accuracy and F1-score are not enough for

Table 5
evaluating predictive models. So another metric which is Overall classification report. (Comparison results between ResNet50,
called the Receiver Op- erating Characteristics curve or ROC InceptionV3,Inception-ResNet, DenseNet, MobileNet, and Xception
curve, is also used for evaluation. With AUC, an model based on average values of Precision, Recall, and F1-score.)
accumulated measure of performance can be defined at Model Recall(%) Precision(%) F1(%)
every possible classification threshold. From the ROC ResNet50 87.00 87.00 87.00
curve, the area under the ROC curve (AUC) is induced, Inception-V3 93.00 93.00 93.00
Inception-ResNet 95.00 95.00 95.00
which is a compatibility indicator of a predictive model.
DenseNet 93.00 93.00 93.00
Derivation of ROC can be done when the True Positive Rate MobileNet 96.00 96.00 96.00
(TPR) is plotted against False Positive Rate (FPR). True Xception 97.00 97.00 97.00
positive rate is nothing but Recall, and FPR is defined by an
Eq. (10)
𝐹𝑃
𝐹𝑃𝑅=
𝐹𝑃 +𝑇𝑁 (10
) In addition, a comparison is established with the other models,
4.5. Results such as ResNet50, InceptionV3, Inception-ResNet, and
DenseNet.
Another compatibility indicator for our proposed models
is ROC
In this segment, we demonstrate the results of our classification results in terms of the number of right
proposed ar- chitectures (MobileNet and Xception) to classifications and misclassification, We presented confusion
scrutinize the robustness of the models. Additionally, the matrices for MobileNet and Xception model in Fig. 9. Fig. 9(a)
experiment was conducted on other deep learning models, illustrates the produced confusion matrix of MobileNet
ResNet50, InceptionV3, Inception-ResNet, and DenseNet, to architectures. From this representation, it can be observed
compare and evaluate the performance of our proposi- tions. that Herpes and Eczema classes have achieved 100% right
Finally, we present the performance comparison of the prediction scores for this approach. The classification
proposed architectures with some graphical presentations performance of Xception architecture is illustrated in Fig.
and tables. 9(b).
A more comprehensive representation of classification
4.5.1. Classification performance of proposed MobileNet and Xception results is depicted in Table 5 of our proposed MobileNet and
models Xception models.
Classification results of our proposed models MobileNet
and Xcep- tion according to our classes (skin disease) are
illustrated in Tables 3 and 4. We have shown the results
based on propositions transfer learning (TL) with
augmentation for each model and without TL and
augmentation. To give an overall insight into our

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
which is presented in Fig. 10. The highest reported micro
average AUC score is 0.9974, which is reported for the
MobileNet model. The lowest micro AUC score is reported
for the ResNet50 model. The ROC of the Xception model is
the second highest, which is 0.9972. Other models also
showed good AUC scores.

4.5.2. Prediction accuracy and loss

In this segment, accuracy and loss for our approaches
are depicted for all six models. In Table 6, validation and
testing accuracy and loss are presented in terms. The
highest testing accuracy is 97.00%, and the lowest loss is
0.16, reported for the Xception model with TL and
augmentation. For MobileNet, the highest accuracy is
96.00%. ResNet50 showed the lowest test accuracy
(86.60%) and highest loss score (2.40) compared to other
models.
In Fig. 11(a) and (b), a line chart is illustrated for
accuracy and loss for the MobileNet and Xception model for
100 epochs. It is seen from Fig. 11(a) that accuracy is pretty
high and consistent for the approach using TL and
augmentation for the MobileNet model. There exist some
reductions and fluctuations per epoch for both models. For
loss, Fig. 11(a) demonstrates that the lowest loss rate is
gained from implementing both TL and augmentation
approaches. For Xception Model, Fig. 11(b) demonstrates
accuracy and loss for each epoch. Like MobileNet, here also
observed high and consistent accuracy scores per epoch by
implementing TL and augmentation. From Fig. 11(b),
insight

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 9. Confusion matrix presenting total number of right and wrong prediction that occurs in the testing process for MobileNet and Xception
model.

Table 7
Total running times for each model.
Model Runtime (s)

ResNet50 14514 s
Inception-V3 5436 s
Inception-ResNet 22971 s
Densenet 16379
MobileNet 7869 s
Xception 10877 s
The effectiveness of our approach in recognizing skin
diseases is depicted in Figs. 12 and 13. we have used both
MobileNet and Xception models with TL and augmentation to
predict diseases as a part of deployment phases. From this
presentation, it can be seen that Both

Fig. 10. ROC curve for deep learning models. This representation depicts the
micro areas under the ROC curve (AUROC) for each of the models.

Table 6
Accuracy and Loss for the best models.
Accuracy (%) Loss
Model
Validation Test Train Validation

ResNet50 96.70 86.60 0.19 2.40

Inception-V3 98.0 93.0 0.05 0.45
Inception-ResNet 98.80 94.80 0.99 0.42
Densenet 94.20 92.80 0.98 0.34
MobileNet (Without TL) 94.45 89.38 0.8 1.77
Xception (Without TL) 95.55 89.79 0.76 1.13
MobileNet (Proposed with TL) 96.00 96.0 0.15 0.21
Xception (Proposed with TL) 97.94 97.0 0.07 0.16

into the loss per epoch can be achieved. Here low loss scores
were reported per epoch by implementing Tl and
augmentation.
Finally, the running time of our training process is given
in Table 7 for each of our models. MobileNet model with TL
and augmentation takes the shortest time (7869 s) to
complete the 100 epochs. The longest time to complete the
execution is reported for the Inception- ResNet model with
the TL approach, 22971 s. The exception model also showed
less time to complete the training process with 10877 s.
We have also presented a comparison between recent
deep learning approaches that are proposed in different
computer vision-based works in Table 8. From this
comparison, it can be clearly derived that our models with
augmentation and transfer learning techniques have better
prediction accuracy.

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

models can accurately predict the classes of the

respective diseases. We can see one misclassification case
for the MobileNet model. But the Xception model can
successfully identify all skin diseases.

4.6. Result analysis

In this research work, we proposed implementing two

deep learning- based architectures MobileNet and Xception,
in recognizing different classes of skin diseases for
computer vision-based applications. Be- sides, other deep
learning models such as ResNet50, InceptionV3, Inception-
ResNet, and DenseNet were also implemented to compare our
approaches’ effectiveness. Finally, we scrutinize the
performance of our different propositions for skin
recognition tasks based on classification reports, confusion
matrix, ROC curves, and classification accuracy.
From Table 3, the decision can be reached about class-
wise classi- fication for both models. The highest Precision
score is 100%, which is achieved for the Nevus class using
both MobileNet and Xception models. Additionally Xception
model also achieved a precision score of 100 for the
Melanoma class. This tells us that our approaches result in a
very good measure of the positive predictions that were
actually correct. For Recall, a maximum of 100 scores is
observed for nevus and Melanoma classes in Xception and
melanoma classes in MobileNet. Since we have used an
imbalanced dataset, F1-score can be a deciding factor. The
maximum score is achieved for Melanoma and Nevus class
using the Xception model. From Table 4, it can be seen that
both Precision and Recall and F1-score are much lower
for cases with no TL and augmentation. We observed the
highest F1-score of 97% for the
Xception (TL+A) model and 96.38% for MobileNet (TL+A)
model. This is an indication that our proposed approach
with TL and augmentation have good classification
capability for imbalanced dataset.
The more comprehensive representation of Precision,
Recall, and F1-score is given in Table 5, where an overall
score for each of the metrics is given for each of the models.
The highest precision is 97.05%, which is reported for the
Xception model. This means that Xception models predict
the correct class of skin disease most of the time. The
highest recall value is 97.00% for Xception and 96.00% for
the MobileNet model, i.e., both of these modes correctly
identify most skin diseases. However, other models such as
ResNet50 performed poorly, achieving a low score. We
observed the highest F1-score of 97.00% for the Xception
model and 96.38% for the MobileNet model.

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Table 8
Comparison between existing approaches and our proposed approaches.
Method/Work done Dataset Used architecture Classification accuracy Best model

Yasir et al. [28] 775 clinical images CNN with adaptive learning 90% CNN
Alarifi et al. [29] Clinical images SVM + CNN 89% CNN with SVM
Li and Shen [30] ISIC 2017 FCRN with LICU 91% FCRN
Rathod et al. [31] DermNet CNN 70% CNN
CNN (PNASNet-5-Large,
Milton [33] ISIC 2018 InceptionResNetV2, 76%,70%, 74%, 67% PNASNet-5-
SENet154, InceptionV4) Large
Liao [34] DermNet and OLE CNN (VGG16) 91% VGG16
(DermNet),
69.5% (OLE)
Shanthi et al. [35] DermNet CNN (ALexNet) 93.3% ALexNet
Kalaiyarivu and Nalini [40] Clinical images CNN 87.5% CNN
Kousis et al. [41] HAM10000 CNN 92.25% DenseNet169
Ahmad et al. [42] Customized CNN + stacked BLSTM 91.73% –
Gupta et al. [39] ISIC VGG16, VGG19, and 82.4%, 83.0%, 83.2% Inception V3
Inception V3 ResNet50,
InceptionV3,
Proposed DermNet + ISIC 2018 Inception-ResNet , 86.60%, 93%, Xception
DenseNet, MobileNet, 94.80%,
and Xception 92.80%, 96%, and
97%

Fig. 11. Accuracy and Loss for MobileNet and Xception model with transfer learning and
augmentation techniques.

Fig. 12. Predicting skin diseases using MobileNet model.

This indicates that our proposed approach with TL and depicted in Fig. 9. Using the transfer learning and
augmentation has good classification capability for augmentation approach, both our models performed very
imbalanced datasets than other models presented in this satisfactorily, outperforming other models. MobileNet and
study. Xception models reported only 20 and 15 misclassification
For illustrating the entire classification and cases, respectively. Accuracy and loss reported by our
misclassification, the confusion matrix as a heatmap is models are presented in Table 6.

1
R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
The highest classification accuracy is 97.00%, which is
observed for the Xception model. The MobileNet model also
gives a tremendous performance with a classification
accuracy of 96.00%. ResNet seems to be a bad choice in
terms of testing accuracy achieving 86.60% testing
accuracy. Both models with TL and augmentation reported
very low loss scores also. But approach with no TL and
augmentation reported a higher loss score with low
accuracy than other approaches.
With the ROC curve presented in Fig. 10, a relation is
established between the false positive rate and the true
positive rate. The highest

1
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)

Fig. 13. Predicting skin diseases using Xception model.

5. Conclusion and future work

This paper has suggested a computer vision-based

approach to recognize five skin diseases Atopic dermatitis,
Eczema, Herpes, Nevus, and Melanoma. Two state-of-the-art
deep learning models MobileNet and Xception, are
implemented to create an automated system. We pro- posed
approaches with transfer learning and augmentation
techniques and evaluated each of the models in computer
vision-based recognition tasks. Augmentation techniques
pave the way for achieving a robust model by increasing the
training data, whereas transfer learning en- ables reusing
the weights from pretrained models. Integrating both
transfer learning and augmentation techniques with
MobileNet and Xception models proved to be a sophisticated
approach to our disease recognition task, outperforming
other deep learning models ResNet50, InceptionV3,
Inception-ResNet, and DenseNet. The MobileNet model has
achieved a classification accuracy of 96.00% and an F1-score
of 96.38%. On the other hand, both test accuracy and F1-
score of the Xception model reached 97.00%. In addition, We
used the Flask web framework to deploy our trained model
by creating a web application that detects skin conditions by
analyzing the skin photograph sup- plied by the client. The
presented approaches can help recognize and diagnose
different dermatological diseases and aid health profession-
als in providing a better healthcare system. Finally, we have
built a web-based framework to identify skin diseases.
For future studies, experiments will be carried out using a
more diverse dataset. Only five classes of skin diseases are
studied. So in the future, we plan to extend our experiment
by adding more classes of skin diseases. Besides, the
approach presented in this paper can be further enhanced by
ensembling different deep learning models. Recently,
transformer-based models such as Vision Transformers
(ViTs) and Mo-
Fig. 14. Developed web interface and application using a skin disease an inference from our saved model.
image.
The web application is created using the Xception model
that has been trained using our skin dataset. Fig. 14 shows
the developed web interface for the clients. The user uploads
ROC–AUC measures are obtained for TL and augmentation the image of a suspected diseased area using any smart
approaches. Finally, Table 7 presents the runtime measures. device and submits it to the developed expert system or the
But when it comes to producing a more robust and accurate web application through the interface. Then feedback will be
classifier, the runtime is a minor fact, whereas accuracy and generated based on our trained model by classifying the
other evaluation metrics are the major ones to consider. image based on different skin conditions.

4.7. Deployment of web application

Finally, we use the Flask [54] web framework to deploy

our trained model. We created a web application that detects
skin conditions by analyzing the skin photograph supplied by
the client. To deploy the flask, we need two routes. First, we
have created an index page route, which will help the users
to upload their images. Finally, a predicted route will create

2
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
bileViT have been widely used in image processing tasks.
So one of our future research directions could be to
develop a transformer-based image recognition model for
skin disease recognition. Our experiments took a huge
computation time, and reducing the computation time in
deep learning approaches could be another potential
research direction of our work.

Declaration of competing interest

The authors declare that they have no known competing

finan- cial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.

References

[1] V. Balaji, S. Suganthi, R. Rajadevi, V.K. Kumar, B.S. Balaji, S.

Pandiyan, Skin disease detection and segmentation using dynamic
graph cut algorithm and classification through Naive Bayes classifier,
Measurement (2020) 107922.
[2] American Cancer Society, Cancer facts & figures for hispanics/latinos
2018–2020, 2020.

2
R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
detecting COVID-19 and pneumonia from chest X-ray images based on
[3] R. Kasmi, K. Mokrani, Classification of malignant melanoma and benign the concatenation of xception and ResNet50V2, Inform. Med. Unlocked
skin lesions: implementation of automatic ABCD rule, IET Image (2020) 100360.
Process. 10 (6) (2016) 448–455, http://dx.doi.org/10.1049/iet- [22] E. Ayan, H.M. Ünver, Diagnosis of pneumonia from chest X-Ray
images using
ipr.2015.0385.
deep learning, in: 2019 Scientific Meeting on Electrical-Electronics &
[4] A.D. Mengistu, D.M. Alemayehu, Computer vision for skin cancer
diagnosis and Biomedical Engineering and Computer Science, EBBT, IEEE, 2019, pp.
recognition using RBF and SOM, Int. J. Image Process. (IJIP) 9 (6) 1–5.
(2015) 311–319. [23] L. Yang, P. Yang, R. Ni, Y. Zhao, Xception-based general forensic
method on
[5] M. Pawar, D.K. Sharma, R. Giri, Multiclass skin disease classification
small-size images, in: Advances in Intelligent Information Hiding and
using neural network, Int. J. Comput. Sci. Inform. Technol. Res. 2 (4)
Multimedia Signal Processing, Springer, 2020, pp. 361–369.
(2014) 189–193.
[24] C. Shi, R. Xia, L. Wang, A novel multi-branch channel expansion network
[6] L.-s. Wei, Q. Gan, T. Ji, Skin disease recognition method based on
for garbage image classification, IEEE Access 8 (2020) 154436–
image color and texture features, Comput. Math. Methods Med.
154452.
2018 (2018).
[25] A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T.
[7] M.N. Islam, J. Gallardo-Alvarado, M. Abu, N.A. Salman, S.P. Weyand, M.
Rengan, S. Said,
Andreetto, H. Adam, MobileNets: Efficient convolutional neural
Skin disease recognition using texture analysis, in: 2017 IEEE 8th
networks for mobile vision applications, 2017, arXiv:1704.04861.
Control and System Graduate Research Colloquium, ICSGRC, 2017, pp.
[26] F. Chollet, Xception: Deep learning with depthwise separable
144–148, http://dx. doi.org/10.1109/ICSGRC.2017.8070584.
convolutions, in:
[8] A. Nawar, N.K. Sabuz, S.M.T. Siddiquee, M. Rabbani, A.A. Biswas, A. Proceedings of the IEEE Conference on Computer Vision and Pattern
Majumder,
Recognition, 2017, pp. 1251–1258.
Skin disease recognition: A machine vision based approach, in: 2021
[27] D.S. Reddy, P. Rajalakshmi, A novel web application framework for
7th International Conference on Advanced Computing and
ubiquitous
Communication Systems, vol. 1, ICACCS, 2021, pp. 1029–1034,
classification of fatty liver using ultrasound images, in: 2019 IEEE 5th
http://dx.doi.org/10.1109/ICACCS51430.
World Forum on Internet of Things (WF-IoT), IEEE, 2019, pp. 502–
2021.9441980.
506.
[9] F. Curia, Features and explainable methods for cytokines analysis
[28] R. Yasir, M.A. Rahman, N. Ahmed, Dermatological disease detection
of Dry Eye Disease in HIV infected patients, Healthc. Anal. 1 using image
(2021) 100001. processing and artificial neural network, in: 8th International
[10] V. Chang, V.R. Bhavani, A.Q. Xu, M. Hossain, An artificial Conference on Electrical and Computer Engineering, IEEE, 2014, pp.
intelligence model
687–690.
for heart disease detection using machine learning algorithms,
Healthc. Anal. 2 (2022) 100016.
[11] S. Dev, H. Wang, C.S. Nwosu, N. Jain, B. Veeravalli, D. John,
A predic-
tive analytics approach for stroke prediction using machine learning
and neural networks, Healthc. Anal. 2 (2022) 100032,
http://dx.doi.org/10.1016/ j.health.2022.100032, URL
https://www.sciencedirect.com/science/article/pii/
S2772442522000090.
[12] R. AlSaad, Q. Malluhi, I. Janahi, S. Boughorbel, Predicting emergency
department
utilization among children with asthma using deep learning models,
Healthc. Anal. 2 (2022) 100050,
http://dx.doi.org/10.1016/j.health.2022.100050, URL
https://www.sciencedirect.com/science/article/pii/S2772442522000181.
[13] M. Ahammed, M.A. Mamun, M.S. Uddin, A machine learning
approach for
skin disease detection and classification using image segmentation,
Healthc. Anal. 2 (2022) 100122,
http://dx.doi.org/10.1016/j.health.2022.100122, URL
https://www.sciencedirect.com/science/article/pii/S2772442522000624.
[14] S. Serte, A. Serener, F. Al-Turjman, Deep learning in medical
imaging: A brief review, Trans. Emerg. Telecommun. Technol. (2020)
e4080.
[15] N.C. Thompson, K. Greenewald, K. Lee, G.F. Manso, The
computational limits of deep learning, 2020, arXiv preprint
arXiv:2007.05558.
[16] H. Pan, Z. Pang, Y. Wang, Y. Wang, L. Chen, A new image
recognition
and classification method combining transfer learning algorithm and
MobileNet model for welding defects, IEEE Access 8 (2020) 119951–
119960.
[17] W. Wang, Y. Li, T. Zou, X. Wang, J. You, Y. Luo, A novel image
classification approach via dense-MobileNet models, Mob. Inf. Syst.
2020 (2020).
[18] K. Sriporn, C.-F. Tsai, C.-E. Tsai, P. Wang, Analyzing lung disease
using highly effective deep learning techniques, Healthcare 8 (2)
(2020) 107.
[19] T. Ghosh, M.M.-H.-Z. Abedin, S.M. Chowdhury, Z. Tasnim, T. Karim,
S.S. Reza,
S. Saika, M.A. Yousuf, Bangla handwritten character recognition using
MobileNet V1 architecture, Bullet. Electr. Eng. Inform. 9 (6) (2020)
2547–2554.
[20] T.M. Angona, A. Siamuzzaman Shaon, K.T.R. Niloy, T. Karim, Z.
Tasnim, S. Reza,
T.N. Mahbub, Automated bangla sign language translation system for
alphabets by means of MobileNet, Telkomnika 18 (3) (2020).
[21] M. Rahimzadeh, A. Attar, A modified deep convolutional neural
network for

2
 R. Sadik, A. Majumder, A.A. Biswas et Healthcare Analytics 3 (2023)
[29] J.S. Alarifi, M. Goyal, A.K. Davison, D. Dancey, R. Khan, M.H. Yap, [49] T.H. Sanford, L. Zhang, S.A. Harmon, J. Sackett, D. Yang, H. Roth,
Z. Xu, D.
Facial skin classification using convolutional neural networks, in:
Kesani, S. Mehralivand, R.H. Baroni, et al., Data augmentation and
International Conference Image Analysis and Recognition, vol. 10317,
transfer learning to improve generalizability of an automated prostate
Springer, Cham, 2017, pp. 479–485, http://dx.doi.org/10.1007/978-3-
segmentation model, Am. J. Roentgenol. (2020) 1–8.
319-59876-5_53.
[50] E. Bisong, Google colaboratory, in: Building Machine Learning and
[30] Y. Li, L. Shen, Skin lesion analysis towards melanoma detection using
Deep Learning Models on Google Cloud Platform, Springer, 2019, pp.
deep learning network, Sensors 18 (2) (2018) 556.
59–64.
[31] J. Rathod, V. Wazhmode, A. Sodha, P. Bhavathankar, Diagnosis of
skin diseases [51] Dermnet, 2020, URL http://www.dermnet.com/. (Accessed 04 November
2020).
using convolutional neural networks, in: 2018 Second International
[52] P. Tschandl, C. Rosendahl, H. Kittler, The HAM10000 dataset, a large
Conference on Electronics, Communication and Aerospace collection
Technology, ICECA, IEEE, 2018, pp. 1048–1051. of multi-source dermatoscopic images of common pigmented skin
[32] M. Chen, P. Zhou, D. Wu, L. Hu, M.M. Hassan, A. Alamri, AI-skin: Skin
lesions, Sci. Data 5 (2018) 180161.
disease recognition based on self-learning and wide data collection [53] N.D. Marom, L. Rokach, A. Shmilovici, Using the confusion matrix for
through a closed-loop framework, Inf. Fusion 54 (2020) 1–9. improving
[33] M.A.A. Milton, Automated skin lesion classification using ensemble ensemble classifiers, in: 2010 IEEE 26-Th Convention of Electrical and
of deep
Electronics Engineers in Israel, IEEE, 2010, pp. 000555–000559.
neural networks in isic 2018: Skin lesion analysis towards melanoma [54] P. Singh, A. Verma, J.S.R. Alex, Disease and pest infection detection in
detection challenge, 2019, arXiv preprint arXiv:1901.10802. coconut
[34] H. Liao, A deep learning approach to universal skin disease tree through deep learning techniques, Comput. Electron. Agric. 182
classification, 2015.
(2021) 105986.
[35] T. Shanthi, R. Sabeenian, R. Anand, Automatic diagnosis of skin
diseases using
convolution neural network, Microprocess. Microsyst. (2020) 103074.
[36] P.N. Srinivasu, J.G. SivaSai, M.F. Ijaz, A.K. Bhoi, W. Kim, J.J. Kang,
Classification
of skin disease using deep learning neural networks with MobileNet
V2 and LSTM, Sensors 21 (8) (2021) 2852.
[37] I. Iqbal, M. Younus, K. Walayat, M.U. Kakar, J. Ma, Automated
multi-class
classification of skin lesions through deep convolutional neural
network with dermoscopic images, Comput. Med. Imaging Graph. 88
(2021) 101843, http://dx.
doi.org/10.1016/j.compmedimag.2020.101843, URL
https://www.sciencedirect.
com/science/article/pii/S0895611120301385.
[38] H.C. Reis, V. Turk, K. Khoshelham, S. Kaya, InSiNet: a deep
convolutional approach to skin cancer detection and segmentation,
Med. Biol. Eng. Comput. 60 (3) (2022) 643–662.
[39] S. Gupta, A. Panwar, K. Mishra, Skin disease classification using
dermoscopy
images through deep feature learning models and machine learning
classifiers, in: IEEE EUROCON 2021 - 19th International Conference
on Smart Technologies, 2021, pp. 170–174,
http://dx.doi.org/10.1109/EUROCON52738.2021.9535552.
[40] M. Kalaiyarivu, N. Nalini, Hand image based skin disease
identification using machine learning and deep learning algorithms,
ECS Trans. 107 (1) (2022) 17381.
[41] I. Kousis, I. Perikos, I. Hatzilygeroudis, M. Virvou, Deep learning
methods for
accurate skin cancer recognition and mobile application,
Electronics 11 (9) (2022) 1294.
[42] B. Ahmad, M. Usama, T. Ahmad, S. Khatoon, C.M. Alam, An ensemble
model of
convolution and recurrent neural network for skin disease
classification, Int. J. Imaging Syst. Technol. 32 (1) (2022) 218–229.
[43] S.F. Aijaz, S.J. Khan, F. Azim, C.S. Shakeel, U. Hassan, Deep learning
application
for effective classification of different types of psoriasis, J. Healthc.
Eng. 2022 (2022).
[44] C.D.S. Duong, Automated fruit recognition using EfficientNet
and MixNet,
Comput. Electron. Agric. 171 (2020) 105326.
[45] A. Khan, A. Sohail, U. Zahoora, A.S. Qureshi, A survey of the
recent architec-
tures of deep convolutional neural networks, Artif. Intell. Rev. 53 (8)
(2020) 5455–5516.
[46] Q. Li, W. Cai, X. Wang, Y. Zhou, D.D. Feng, M. Chen, Medical image
classification
with convolutional neural network, in: 2014 13th International
Conference on Control Automation Robotics & Vision, ICARCV, IEEE,
2014, pp. 844–848.
[47] O. Ukwandu, H. Hindy, E. Ukwandu, An evaluation of lightweight
deep learning
techniques in medical imaging for high precision COVID-19
diagnostics, Healthc. Anal. 2 (2022) 100096,
http://dx.doi.org/10.1016/j.health.2022.100096, URL
https://www.sciencedirect.com/science/article/pii/S2772442522000417
.
[48] K. Guzel, G. Bilgin, Classification of breast cancer images using
ensembles of transfer learning, Sakarya Üniv. Bilimleri EnstitÜSÜ
Dergisi 24 (5) (2020) 791–802.