Classification of Human Monkeypox Disease Using

Deep Learning Models and Attention Mechanisms

Md. Enamul Haque Md. Rayhan Ahmed
Department of Computer Science and Engineering Department of Computer Science and Engineering
United International University United International University
mhaque182055@bscse.uiu.ac.bd rayhan@cse.uiu.ac.bd

Razia Sultana Nila Salekul Islam

Department of Computer Science and Engineering Department of Computer Science and Engineering
United International University United International University
rnila182041@bscse.uiu.ac.bd salekul@cse.uiu.ac.bd

Abstract—As the world is still trying to rebuild from the calls for keeping isolation and a distance from other people,
destruction caused by the widespread reach of the COVID-19 and so on.
virus, and the recent alarming surge of human monkeypox
outbreaks in numerous countries threatens to become a In recent research, it is clear that deep learning is far more
new global pandemic too. Human monkeypox disease symptoms sophisticated than classical machine learning models[3] like
are quite similar to chickenpox, and measles classic symptoms, KNN, Random Forest classifier, and so on. However, they
with very intricate differences such as skin blisters, which come remain machines that solve the categorization problem by
in diverse forms. Various deep-learning methods have shown requiring additional domain knowledge and human
promising performance in the image-based diagnosis of Covid- involvement. Additionally, traditional machine learning
19, tumor cell, and skin disease classification tasks. In this models are typically only effective for the tasks for which they
paper, we try to integrate deep transfer learning-based methods, were created. Therefore, the traditional machine learning
along with a convolutional block attention module (CBAM) to approach may not be as effective in this subject.
focus on the relevant portion of feature maps to conduct an
image-based classification of human monkeypox disease. We Due to the superior learning capabilities of deep learning
implement five deep learning models—VGG19, Xception, models notably the variants of convolutional neural networks
DenseNet121, EfficientNetB3, and MobileNetV2 along with (CNNs), have recently transformed various fields of medical
integrated channel and spatial attention mechanisms and science [4][5]. These deep networks can analyze images in
perform a comparative analysis among them. An architecture several layers, automatically identify significant features, and
consisting of Xception-CBAM-Dense layers performed better learn to discover the best representations for particular tasks
than the other models at classifying monkeypox and other when trained with a huge amount of data. Also transferring
diseases with a validation accuracy of 83.89%. learning models helps to tackle the problem of the enormous
dataset demand.
Keywords—Monkeypox, Classification, Image-based
diagnosis, Deep Learning, Channel attention, Spatial Attention. Transfer learning is another method that is frequently
employed when there is a lack of data. With this technique, a
I. INTRODUCTION CNN model that has already been trained on a big dataset
According to World Health Organization (WHO) [1], (such as ImageNet) is used to transfer its knowledge to a
Monkeypox is a rare disease caused by the monkeypox virus. separate, much smaller dataset for context-specific learning.
It can spread from animals to people and is transferable As a result, employing the transfer learning method will speed
between people and the environment since it is a viral zoonotic up, simplify, and improve our performance compared to other
infection. The monkeypox outbreak became the most talked- traditional ways.
about topic in the world after the COVID-19 pandemic was In this study, we used the Monkeypox Skin Lesion Dataset
over. Up to this point, a total of 72,874 confirmed cases have (MSLD) [6] and a few transfer-learning-based models with
been discovered, including 28 deaths across 12 locations[2]. the CBAM attention mechanism. Used transfer learning
Fever, headache, muscle aches, back discomfort, low energy, model for comparing is - VGG19 [7], DenseNet121[8],
and swollen lymph nodes are the most typical signs of Xception [9], EfficientNetB3 [10] and MobileNetV2 [11] and
monkeypox, which can also be followed by the appearance of we were able to predict the enhanced train data with an
a rash[1]. The face, palms of the hands, soles of the feet, groin, accuracy of 69.86% for VGG19 and 78.27% for
genital, and/or anal regions might all be affected by the DenseNet121, 54.21% for EfficientNetB3 model, 74.07% for
rash[1]. The slight distinction between monkeypox and Others MobileNetV2 model and 79.90% for proposed Xception
(measles, chickenpox) is inflammation and rash on the body, model.
which is difficult to detect by human eyes except by the
Polymerase Chain Reaction (PCR) Test. The remaining papers are divided into the following
sections, with section II covering the specifics of the related
It is exceedingly difficult for medical professionals to make work with our paper, section III states the methodology and
an accurate diagnosis of this disease quickly. Additionally, it related materials that are required for the study, section IV
is difficult to find the PCR test used to detect the monkeypox states experimental details like dataset description,
virus. However, prompt diagnosis of this illness is necessary hyperparameter tuning information and evaluation metrics
to stop the spread by adhering to other WHO-provided details and with result analysis.
guidelines and the first step of preventing monkeypox, which
 II. RELATED WORK feature maps with we incorporate the convolutional block
Recently many researchers have explored diverse deep attention module.
learning algorithms for the detection of human monkeypox
disease. Ali et al. [6] Incorporated VGG-16, ResNet50, and
InceptionV3 pre-trained models for the classification of
monkeypox and other diseases (chickenpox, measles) on the
Monkeypox Skin Lesion Dataset (MSLD) which was
developed by the authors. The dataset contains images of skin
lesions in two categories. One is Monkeypox and Others
(Chickenpox, Measles). The overall accuracy ResNet50
achieved is 82.96(±4.57%), and VGG-16 achieved
81.48(±6.87%). The author ensembled these three models
and resulting in an accuracy of 79.26 (±1.05%).

In another study, Islam et al. [12] examined the

performance of the monkeypox classification over pre-
trained ResNet50, DenseNet121, Inception-V3, SqueezeNet,
MnasNet-A1, MobileNet-V2, and ShuffleNet-V2 models.
Based on ShuffleNet-V2 model obtained a maximum of 79%
accuracy in comparison to the other models used, with 85%
precision, demonstrating the possibility of employing AI
models. The Monkeypox Skin Image Dataset 2022 was also
a dataset the authors developed and used
Ahsan et al. [13] created the Monekypox2022 image
dataset first before proposing a modified VGG16 model using
two separate studies. The suggested modified VGG16 has an
accuracy of 97±1.8% (AUC: 97.2) for study one and
88±0.8% (AUC: 0.867) for study two in identifying
monkeypox patients. Additionally, they talked about their
model's feature extraction and prediction using Local
Interpretable Model-Agnostic Explanations (LIME).
Sitaula et al. [14] analyzed and evaluated thirteen different
pre-trained deep-learning models for the detection of the
monkeypox virus. DenseNet-169 and Xception performed
better than others on the Monkeypox2022 image dataset.
Based on performance DenseNet-169 and Xception were
ensembled. Additionally, the ensembled method offers
accuracy (87.13%), precision (85.44%), recall (85.47%), and
F1-score (85.40%).
Sahin et al. [15] used some pre-trained deep learning
models on the MSLD dataset and discovered that
MobileNetV2 and EfficientNetB0 model performed well. Figure 1. Proposed Architecture
They then constructed a mobile application that can
categorize monkeypox disease after converting the entire
model into a TensorFlow lite model embedding with At first Human Monkeypox Skin Lesion Dataset have been
metadata. prepared for training by resizing the image to a resolution of
224 x 224 x 3. To be noted, we have frozen every layer except
III. METHODOLOGY the last two layers of all the pre-trained architectures. To
To test the feasibility of classifying human monkeypox in utilize the most relevant parts of the generated feature maps
this analysis, we picked five pre-trained deep learning models we utilized the convolutional block attention module
through the transfer learning process, ranging from heavy- (CBAM), proposed by Woo et al. [16]. CBAM contains two
weight architectures like VGG-19 and Xception to sequential attention-based mechanisms - Channel attention
lightweight models like DenseNet121, MobileNetV2, and followed by Spatial attention. The taken output from the
EffcientNetB3. Figure 1 depicts the overall pipeline of the CBAM was passed through two dense layers of 256 and 128
training procedure for those models. Due to the lack of neurons with ReLU activation function. For the conversion
images in the utilized dataset, we selected deep models over of the layer into 1D to feed the last layer, we used Flatten and
a variety of sizes to examine the impact of training sample fed the output into the last dense layer with sigmoid activation
size. All pre-trained models are customized using the same to perform binary classification. Finally, validation data were
method. Moreover, to focus the network on more relevant used to synthesize the outcome from the training data.
 CBAM is an adaptive image refinement module, which is process applies average pooling and maximum pooling,
sequentially applied in the channel and spatial dimensions which are then concatenated to produce the feature
and is used to concentrate more on major features of the descriptor. After that, a convolutional layer was used to create
images. The overall summarized attention process is: the Ms(F) ∈ RH×W spatial attention map. A conventional
convolutional layer concatenates the results of the pooling
F ′ = Mc (F) ⊗ F ………………… (1) operations, Fsavg ∈ R1xHxW and Fsmax ∈ R1xHxW to create a 2D
spatial attention map. The calculations are defined as follows:
F ′′ = Ms (F ′) ⊗ F ′ ……………… (2)
Ms(F) = σ (f 7×7 ([AvgPool(F); MaxPool(F)]))
CxHxW
Where F = feature map (F ∈ R = input), 1D channel
attention map Mc ∈ R C×1×1 and 2D spatial attention map Ms = σ (f 7×7 ([Fsavg; Fsmax])) ……………… (4)
∈ R 1×H×W and F′′ is the final refined output, and ⊗ denotes
element-wise multiplication. Here, σ = Sigmoid function, f 7×7 = Convolutional operation,
and filter size of 7x7.

IV. EXPERIMENTS AND ANALYSIS

A) Dataset
To classify Monkeypox among chickenpox, measles, and
monkeypox we used Monkeypox Skin Lesion Dataset
(MSLD) [6]. This image dataset includes a total of 228
original images of measles and chickenpox. Some of the
Figure 2. Design of the Channel attention module [16].
images are demonstrated in fig. 3 and fig. 4. The dataset
includes both original images and augmented folded images
Channel attention [16] as illustrated in Figure 2. focuses (Train, Test, Validation) with a proportion of 70:10:20. We
on the inter-channel relationship of features and the reduction used the data from the augmented Train folder for training,
of channel redundancy. The features of average-pooling and which contains 2142 images in 2 classes (Monkeypox,
max-pooling were combined to aggregate the two separate Others). Here, the training and validation images from the
spatial descriptors Fcavg and Fcmax. dataset were augmented while the test set contained only the
original images.
The channel attention map Mc ∈ RCx1x1 is obtained by feeding
these descriptors into a shared network made up of a multi-
layer perceptron (MLP) with a single hidden layer, after
which the output feature vectors are subjected to a merging
process. Overall, the calculation is:

Mc(F) = σ (MLP (AvgPool(F)) + MLP (MaxPool(F)))

= σ (W1(W0(Fcavg)) + W1(W0(Fcmax))) ……….…… (3) Figure 4. Monkeypox skin lesion sample images.

Here, σ = sigmoid function, W0, W1 = MLP weights.

Figure 5. Others (chickenpox, measles) skin lesion sample

images.

B) Optimization and hyper-parameter tuning

Figure 3. Design of the spatial attention module [16] . In this paper, we have used various hyper-parameter to ensure
a good performance from the model. In order to reduce the
losses, we used adam optimizer with a learning rate of 0.001.
Spatial attention [16] as demonstrated in Figure 3. focus We also employed binary_crossentropy that computes the
on the informative part or content information of an input cross-entropy loss between true labels and predicted labels
feature in the spatial position. On the input feature, the
with accuracy metrics too. To get the best model from the From Table I. we can observe that the Xception-CBAM-
execution we have employed Model Checkpoint with other Dense architecture performed better than other investigated
required tunings for monitoring val_loss metrics to monitor combinations with a validation accuracy, precision, recall and
the behavior of the accuracy of the model. For all f1-score of 83.89%, 90.70%, 89.10%, and 90.11%,
experiments, the batch size was fixed at 32. respectively. All the scores presented in Table I are validation
scores. As for the value of evaluated metrics on the test
C) Evaluation Metrics dataset, the VGG19-CBAM-Dense architecture, Xception-
CBAM-Dense architecture, DenseNet121-CBAM-Dense
The most prominent statistical methods, including accuracy, architecture, MobileNetV2-CBAM-Dense architecture, and
precision, recall, and F1-score, are employed to measure and lastly the EfficientNetB3-CBAM-Dense architecture
present the overall experimental outcome. They are defined achieved a test accuracy of 70.71%, 85.83%, 79.28%,
as follows: 75.28%, and 84.37%, respectively.
TP  TN …………………… (5)
Accuracy 
TP  TN  FP  FN TABLE I: A comparative summary of the acquired
TP evaluation metric scores (validation) using different variants
Precision  ………………………………... (6) of the pre-trained models along with CBAM and Dense
TP  FP
layers, and with other studies that used the same MSLD
TP …………………………………… (7)
Recall  dataset (- denotes not mentioned).
TN  FP
Precision*Recall ………………….. (8) Model Accuracy Precision Recall F1_Score
F1-Score=2* Architecture (%) (%) (%) (%)
Precision+Recall VGG19-CBAM- 71.86 73.91 72.90 73.31
Dense
D) Result Analysis Xception- 83.89 90.70 89.10 90.11
CBAM-Dense
For the classification of monkeypox among other similar DenseNet121- 78.27 80.49 76.31 81.56
diseases (Chickenpox, Measles) we have employed five pre- CBAM-Dense
trained architecture models with attention mechanisms and MobileNetV2- 74.07 77.16 72.22 75.19
CBAM-Dense
with dense layers. It should be noted that the results presented
EfficientNetB3- 81.43 85.61 81.05 86.38
in Table I are the average of performance across four different CBAM-Dense
folds (4-fold cross-validation). The confusion matrix Ensemble [6] 79.26 84.00 79.00 81.00
generated each of the four-fold is presented in fig. 6. We have ShuffleNet [15] 80.00 - - -
stated the average accuracy, precision, recall, and f1-score of
the VGG19, DenseNet121, EfficientNetB3, MobileNetV2,
and Xception architectures along with CBAM and Dense V. CONCLUSION
layers for classification in Table I. This research work presents a comparative analysis and
preliminary feasibility study of five pre-trained deep learning
algorithms along with a convolutional block attention module
(CBAM) and dense layers for the classification of
monkeypox disease from skin lesions by using the
Monkeypox Skin Lesion Dataset (MSLD). Implementation
of the CBAM module with channel and spatial attention
module permits the proposed network to emphasize the major
and effective feature maps from the image features and focus
on inter-channel dependencies of the diseased regions of the
image. The proposed model obtains reasonable classification
performance with a validation accuracy of 83.89% using the
(a) Fold 1 (b) Fold 2 Xception-CBAM-Dense layer architecture. The proposed
study produced superior findings when compared to other
studies using the same dataset. One of the major drawbacks
to the proper training of the proposed model is the extremely
small number of original images of monkeypox and other
diseases in the dataset. In future work, we hope to design a
more optimized model and train it on a larger dataset of
human monkeypox disease images.

(c) Fold 3 (c) Fold 4 REFERENCES

Figure 6. Generated confusion matrix from the 4-fold cross [1] “Monkeypox: Questions and Answers,”
validation. https://www.who.int/news-room/questions-and-
answers/item/monkeypox?gclid=CjwKCAjwg5uZBhATEiwAh
hRLHkkxZ8te7FJSn9IVwo_uWb3bua5tjv0gtxwibxbZqzwJ2xo
mhOWJbxoCg48QAvD_BwE, Aug. 31, 2022.
 [2] “2022 Monkeypox Outbreak Global Map | Monkeypox |
Poxvirus CDC.”
https://www.cdc.gov/poxvirus/monkeypox/response/2022/worl
d-map.html (accessed Oct. 14, 2022).
[3] I.H. Sarker, “Deep Learning: A Comprehensive Overview on
Techniques, Taxonomy, Applications and Research Directions,”
SN Comput Sci, vol. 2, no. 6, pp. 1–20, Nov. 2021, doi:
10.1007/S42979-021-00815-1/FIGURES/13.
[4] D. Ravi et al., “Deep Learning for Health Informatics,” IEEE J
Biomed Health Inform, vol. 21, no. 1, pp. 4–21, Jan. 2017, doi:
10.1109/JBHI.2016.2636665.
[5] F. Wang, L. P. Casalino, and D. Khullar, “Deep Learning in
Medicine-Promise, Progress, and Challenges,” JAMA Intern
Med, vol. 179, no. 3, pp. 293–294, Mar. 2019, doi:
10.1001/JAMAINTERNMED.2018.7117
[6] S. N. Ali et al., “Monkeypox Skin Lesion Detection Using Deep
Learning Models: A Feasibility Study,” Jul. 2022, [Online].
Available: http://arxiv.org/abs/2207.03342
[7] K. Simonyan and A. Zisserman, “Very Deep Convolutional
Networks for Large-Scale Image Recognition,” 3rd International
Conference on Learning Representations, ICLR 2015 -
Conference Track Proceedings, Sep. 2014, doi:
10.48550/arxiv.1409.1556.
[8] G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger,
“Densely Connected Convolutional Networks,” Proceedings -
30th IEEE Conference on Computer Vision and Pattern
Recognition, CVPR 2017, vol. 2017-January, pp. 2261–2269,
Aug. 2016, doi: 10.48550/arxiv.1608.06993.
[9] F. Chollet, “Xception: Deep Learning with Depthwise Separable
Convolutions,” Proceedings - 30th IEEE Conference on
Computer Vision and Pattern Recognition, CVPR 2017, vol.
2017-January, pp. 1800–1807, Oct. 2016, doi:
10.48550/arxiv.1610.02357.
[10] M. Tan and Q. v. Le, “EfficientNet: Rethinking Model Scaling
for Convolutional Neural Networks,” 36th International
Conference on Machine Learning, ICML 2019, vol. 2019-June,
pp. 10691–10700, May 2019, doi: 10.48550/arxiv.1905.11946.
[11] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. C. Chen,
“MobileNetV2: Inverted Residuals and Linear Bottlenecks,”
Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, pp. 4510–4520, Jan.
2018, doi: 10.48550/arxiv.1801.04381
[12] T. Islam, M. A. Hussain, F. Uddin, H. Chowdhury, and B. M. R.
Islam, “Can Artificial Intelligence Detect Monkeypox from
Digital Skin Images?”, doi: 10.1101/2022.08.08.503193.
[13] M. M. Ahsan, M. R. Uddin, M. Farjana, A. N. Sakib, K. al
Momin, and S. A. Luna, “Image Data collection and
implementation of deep learning-based model in detecting
Monkeypox disease using modified VGG16,” Jun. 2022,
[Online]. Available: http://arxiv.org/abs/2206.01862
[14] C. Sitaula and T. B. Shahi, “Monkeypox virus detection using
pre-trained deep learning-based approaches,” Sep. 2022,
[Online]. Available: http://arxiv.org/abs/2209.04444
V.[15] H. Sahin, I. Oztel, and G. Yolcu Oztel, “Human Monkeypox
Classification from Skin Lesion Images with Deep Pre-trained
Network using Mobile Application,” Journal of Medical
Systems 2022 46:11, vol. 46, no. 11, pp. 1–10, Oct. 2022, doi:
10.1007/S10916-022-01863-7.
[16] S. Woo, J. Park, J. Y. Lee, and I. S. Kweon, “CBAM:
Convolutional Block Attention Module,” Lecture Notes in
Computer Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics), vol.
11211 LNCS, pp. 3–19, Jul. 2018, doi:
10.48550/arxiv.1807.06521.