Skip to main content
NCBI home page
Scientific Reports logoLink to Scientific Reports
. 2024 Nov 4;14:26579. doi: 10.1038/s41598-024-75030-y

Leveraging convolutional neural networks and hashing techniques for the secure classification of monkeypox disease

Essam Abdellatef 1, Alshimaa H Ismail 2,✉, M I Fath Allah 3, Wafaa A Shalaby 4
PMCID: PMC11535243  PMID: 39496621

Abstract

The World Health Organization declared a state of emergency in 2022 because of monkeypox. This disease has raised international concern as it has spread beyond Africa, where it is endemic. The global community has shown attention and solidarity in combating this disease as its daily increase becomes evident. Various skin symptoms appear in people infected with this disease, which can spread easily, especially in a polluted environment. It is difficult to diagnose monkeypox in its early stages because of its similarity with the symptoms of other diseases such as chicken pox and measles. Recently, computer-aided classification methods such as deep learning and machine learning within artificial intelligence have been employed to detect various diseases, including COVID-19, tumor cells, and Monkeypox, in a short period and with high accuracy. In this study, we propose the CanDark model, an end-to-end deep-learning model that incorporates cancelable biometrics for diagnosing Monkeypox. CanDark stands for cancelable DarkNet-53, which means that DarkNet-53 CNN is utilized for extracting deep features from Monkeypox skin images. Then a cancelable method is applied to these features to protect patient information. Various cancelable techniques have been evaluated, such as bio-hashing, multilayer perceptron (MLP) hashing, index-of-maximum Gaussian random projection-based hashing (IoM-GRP), and index-of-maximum uniformly random permutation-based hashing (IoM-URP). The proposed approach’s performance is evaluated using various assessment issues such as accuracy, specificity, precision, recall, and fscore. Using the IoM-URP, the CanDark model is superior to other state-of-the-art Monkeypox diagnostic techniques. The proposed framework achieved an accuracy of 98.81%, a specificity of 98.73%, a precision of 98.9%, a recall of 97.02%, and fscore of 97.95%.

Keywords: Monkeypox, Cancelable techniques, CNN, And DarkNet-53

Subject terms: Drug discovery, Diseases, Health care, Health occupations, Medical research

Introduction

The Monkeypox virus is a smallpox virus, an infectious disease related to the zoonotic virus Orthpoxvirus, which is related to cowpox1. Monkeypox cases have been increasing in Africa and other countries2. This disease is transmitted by rodents and monkeys and easily spreads from one human to another during body contact. In 1958, the first discovery of this disease by Preben Alexander in monkeys was in a laboratory in Denmark3. The first detection of this disease in humans was recorded in 1970 in the Congo4. Subsequently, the virus spread to other countries, such as Liberia, Sierra Leone, and Nigeria5. In 2003, the first case of Monkeypox appeared in the United States and was the result of contact with prairie dogs that were sick without being reported. From 2017 to 2019, the disease began to spread in Nigeria, with 122 deaths reaching 7 cases6. To date, the number of Monkeypox cases has reached 90,000, and 115 deaths have been reported. It has spread widely to countries outside Africa, such as the United States of America, Germany, France, Australia, and Sweden7–16.

The main symptoms that appear early in cases influenced by this disease are headache, fever, chills, muscle pain, fatigue, and skin rashes, such as chickenpox17. Monkeypox is considered a viral disease similar to coronavirus, but it is less widespread. Despite this, Monkeypox began to spread in Africa in 1990, until cases increased and began to spread until the number of cases reached 5,000 in 2020, after which it began to spread in the United States of America and Europe18. According to the Centers for Disease Prevention, the number of cases reached approximately 83,424 by 2022. There is still no cure for the Monkeypox virus, according to what was praised by the National Center for Disease Eradication. However, vaccination is carried out using two drugs, ticovirimat and brincidofovir, which are used to treat the smallpox virus, according to what was provided by the Center for Disease Control (CDC)19. This vaccination has become a solution to prevent the Monkeypox virus in many countries, but the United States of America no longer gives it to humans20. Procedures for diagnosing Monkeypox include initial observations of the unusual characteristics of existing skin lesions and a history of current exposure. This virus can be diagnosed using an electron microscope to test for skin diseases, and it can also be diagnosed using PCR, as it has also been used to diagnose COVID-19 patients21. As of September 27, 2023, approximately 90,630 cases of Monkeypox had been confirmed in laboratories, with 115 deaths recorded. The following figure shows the confirmed cases of the top 15 countries in terms of the number of confirmed infections, as the United States of America is considered the highest country, with the rate of infections reaching 30,861 and deaths 54, followed by Brazil, where the number of infections reached 10,967 and deaths 16. Furthermore, the number of confirmed cases in Spain reached 7,850, with three deaths. Colombia, France, Germany, and Mexico have approximately 4000 cases of this disease, while Argentina, China, Chile, Canada, and Portugal have around 1000 cases19.

Deep learning (DL) is a subdivision of machine learning (ML) that takes cues from the operation of the human brain. It is used to extract and categorize characteristics in images. The primary advantage of this technique is its use in unsupervised learning, where learning is done from unlabeled data examples. Due to its ability to consume unlabeled data, operate without feature engineering, and make highly accurate and precise predictions, machine learning has become a widely used technology in industries such as self-driving cars, face recognition, object detection, and image classification. Recent research has proven that computer vision, machine learning, and deep learning can be utilized to automate the identification of various illnesses in the human body22. Feature extraction stage is conducted using a convolutional neural network (CNN), which is a subset of deep learning. It is in charge of acquiring an input image and allocating learn-able weights as well as biases to different objects in that image. It utilizes appropriate filters to extract temporal and spatial dependencies within an image23. Each layer in the CNN carries out a distinct operation on the input images. A variety of innovations has led to a notable enhancement in the representational capacity of CNN. Notably, there has been significant attention given to the idea of leveraging the information, depth, and width of the architecture. Similarly, the concept of utilizing a cluster of layers as a structural component is becoming increasingly popular. A Convolutional Neural Network (CNN) is composed of various layers, each with a distinct function in processing input images. The key layers that contribute to constructing CNN architectures include convolutional, pooling, batch normalization, depth concatenation, and fully connected layers. The main contributions of this study are as follows:

  1. Providing a secure Monkeypox classification system to sufficiently diagnosis, Monkeypox cases and impose complete confidentiality about patient information.

  2. Extraction of appropriate and distinctive features from Monkeypox skin images using a highly efficient pre-trained CNN (DarkNet-53).

  3. Preserving patient information using cancelable techniques.

In the next sections, we discuss the related work in Sect. 2, the architecture of the proposed CanDark model in Sect. 3, the experimental results in Sect. 4. Finally, the paper is concluded in Sect. 5.

Related work

There are various advanced techniques for identifying and categorizing illnesses through deep learning models, including the identification of COVID-19. Ozkaya et al.24 proposed a hybrid model for classification that employed an SVM classifier utilizing deep feature extraction approaches such as Resnet-50, GoogleNet, and VGG-16. Cohen et al.25 established a strategy for predicting the severity score of COVID-19 pneumonia using X-ray pictures. Various pre-trained networks were utilized to categorize Monkeypox infections in26,27. More than a million images were used to train the pre-trained networks. The task they have is to categorize input images into 1000 or more element classes, such as pencil, coffee mug, keyboard, and many others. The utilization of these networks, in addition to transfer learning, provides easier and faster performance than the training of the model from scratch. The concept of deep transfer learning depends on using these pre-trained networks to classify new input images28,29. Table 1 provides previous and relevant research on various Monkeypox classification approaches.

Table 1.

Previous and relevant research.

Method Dataset Classes Results
Sahin (2022)30 228 images (MSLD) Binary

The accuracy of

ResNet18 = 86.8%,

GoogleNet = 82.22%, EfficientNetbo = 91.11%, NasnetMobile = 86.6%, ShuffleNet = 80%, MobileNetv2 = 91.11%
Irmak (2022)31 770 images in MSLD Multiclass

Accuracy = 91.38%

Precision = 90%

Recall = 86%

Fscore = 88%
Alcala (2023)32 2067images of (MSLD) Binary

Accuracy = 97%

AUC = 0.76

Loss function = 0.1442
Akin (2022)33 572 images in Monkeypox Skin Images Dataset (MSID) Binary

Accuracy = 98.25%,

Sensitivity = 96%

Fscore = 98%
Sitaula (2022)34 1753 images in Monkeypox dataset Multiclass

Precision = 85%

Recall = 85%

Fscore = 85%

Accuracy = 87.13%
Ali (2022)35 (MSLD) which was developed by the authors.

Binary

Monkeypox and Others (Chickenpox, Measles).

The accuracy of

ResNet50 = 82.96 (± 4.57%), VGG16 = 81.48 (± 6.87%).
Altun (2023)36 2056 images with augmentation of Monkeypox Skin Lesion Dataset. Two different classes (positive and negative)

Fscore = 98%

Accuracy = 96%

Recall = 97%
Islam (2022)37 Monkeypox Skin Image Dataset 2022 Binary ShuffleNet-V2 model has achieved the best results of 79% accuracy with 85% precision.
Haque (2022)38 572 images of MSID Binary Xception-CBAM-Dense accuracy = 83.89%
Ahsan (2022)39

90 images for study one and 1754 images for study two of “Monkeypox 2022”

Which was collected by the authors

 (Chickenpox, Measles, and Normal images) images

(study one)

Accuracy = 97 ± 1.8%

AUC = 0.97

(study two)

Accuracy = 88 ± 0.8%

AUC = 0.867
Open in a new tab

The proposed Framework

To preserve patient information and impose complete confidentiality, we need a secure classification system. The proposed approach is a secure machine-learning-based system for classifying various cases of Monkeypox disease (Fig. 1). The proposed approach incorporates machine learning to extract sufficient, appropriate, and distinctive features from input images. The DarkNet-53 CNN model is developed and used for feature extraction. To protect patient information, various cancelable techniques are evaluated: Bio-Hashing, MLP Hashing, IoM-GRP, and IoM-URP. Cancelable methods are applied to the extracted features before the classification stage. Cancelable methods are preferred to other encryption techniques as they provide efficient protection of information and, at the same time, do not affect the classification accuracy of the system. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Fig. 1.

Fig. 1

Open in a new tab

The proposed CanDark model.

Dataset

The Monkeypox Skin Image Dataset is used, which contains 770 images classified into four categories: (107) chickenpox cases, (91) measles cases, (279) cases of Monkeypox (279), and (293) normal cases40. Classes are divided into 70% for training and 30% for testing. Table 2 describes the Monkeypox Skin Image Dataset. Furthermore, Fig. 2 provides a sample of images for all classes.

Table 2.

Number of images, training set, and testing set for the classes (Chickenpox, Measles, Monkeypox, and normal) of the Monkeypox skin image dataset.

Class #Images Training set Testing set
Chickenpox 107 74 33
Measles 91 64 27
Monkeypox 279 196 83
Normal 293 205 88
Open in a new tab

Fig. 2.

Fig. 2

Open in a new tab

Sample of images of the classes (Chickenpox, Measles, Monkeypox, and Normal) of the Monkeypox Skin Image Dataset.

Feature extraction network: DarkNet-53

DarkNet-53 is a convolutional neural network that extracts features from images. This network consists of 53 deep convolution layers followed by a batch normalization (BN) layer and a Leaky ReLU layer which represents the activation function. Leaky ReLU layer could be described as the following: Inline graphic Leaky ReLU layer provides a small slope for negative values. Furthermore, DarkNet-53 is the backbone of the YOLOv3 object detection methodology. The architecture of the DarkNet-53 is described in details in Table 3.

Table 3.

Architecture of feature extraction network (DarkNet-53).

Type Filters Size Output
Convolutional 32 3 × 3 256 × 256
Convolutional 64 3 × 3 /2 128 × 128
1× Convolutional 32 1 × 1 128 × 128
Convolutional 64 3 × 3
Residual
Convolutional 128 3 × 3 /2 64 × 64
2× Convolutional 64 1 × 1 64 × 64
Convolutional 128 3 × 3
Residual
Convolutional 256 3 × 3 /2 32 × 32
8× Convolutional 128 1 × 1 32 × 32
Convolutional 256 3 × 3
Residual
Convolutional 512 3 × 3 /2 16 × 16
8× Convolutional 256 1 × 1 16 × 16
Convolutional 512 3 × 3
Residual
Convolutional 1024 3 × 3 /2 8 × 8
4× Convolutional 512 1 × 1 8 × 8
Convolutional 1024 3 × 3
Residual
Average Pooling
Fully Connected 1000
Softmax -
Open in a new tab

Cancelable biometric techniques

Cancelable biometrics is the best protection system from the point of view of researchers and is now used to enhance template security and privacy protection in biometric systems41. Protection is achieved by converting the original properties into another property with a new representation that is not reversible using certain transformations. The new form after the conversion is called a revocable biometric template. A cancelable biometric must possess four characteristics: irreversibility, diversity, reversibility, and reusability42. There are several cancelable biometric techniques, and we concentrate on image-hashing techniques. One of the basic functions of encryption is a hash function. This is called the compression or contraction function and can be used in passwords and file authentication. The hash takes a number of bits from the input and outputs, and 128 bits is a fixed number because it is impossible to recover the original from the output43. Password authentication requires a comparison of the stored copies of the hash with the entered password hash.

Bio-hashing

Bio-Hashing is considered as a feature extraction technique that uses wavelet transform to extract biometric features44. The features x, Inline graphic are extracted from input data for N bits. Also, the orthogonal random vector is defined as Inline graphic The dot product of the feature vector and the random vector is calculated to represent the Bio-Hash C as in the following equation:

graphic file with name M4.gif 1

where sig is denoted as the signum function and Inline graphic is the threshold. Additionally, the operation of Bio-Hashing technique is described through the following diagram in Fig. 3.

Fig. 3.

Fig. 3

Open in a new tab

Bio-hashing representation.

Multi-layer perceptron (MLP) hashing

Using MLP hashing offers many useful attributes. Changing the values of the initial weights of the network leads to a complete difference in the segmentation output. Also, the construction of the network can be easily modified by adding new neurons to generate many output bits. MLP network is used as one-way hashing. MLP Network consists of several layers of neurons that called input (i), hidden (o) and output (y) layers as described in Fig. 4. The output layers are from y0 to y127, O0 to O63 are the output of the hidden layers, and the weights are denoted as Wx, y45. The sigmoidal function is used for finding the o and I. The range of the output is between 0 and 1.

Fig. 4.

Fig. 4

Open in a new tab

MLP layers and representation in a matrix.

Index-of-maximum Gaussian random projection-based hashing (IoM-GRP)

The IoM hashing type converts the biometric feature vector to maximum ranking hashing code. IoM-GRP is derived from IoM hashing and depends on Gaussian random projection vector46. IoM-GRP could be summarized in the following diagram in Fig. 5.

Fig. 5.

Fig. 5

Open in a new tab

IoM-GRP algorithm.

Index-of-maximum uniformly random permutation-based hashing, (IoM-URP)

Another type of IoM hashing is IoM-URP. The IoM-URP hashing method relies on random permutation46. Weights are controlled through choosing the variable k of the inputs of the biometric vectors. IoM-URP algorithm is depicted in Fig. 6.

Fig. 6.

Fig. 6

Open in a new tab

IoM-URP algorithm.

The hashing techniques chosen; Bio-Hashing, MLP Hashing, IoM-GRP, and IoM-URP present unique benefits compared to alternative hashing methods. These techniques enhance security through non-invertible transformations, ensuring that original biometric data remains unrecoverable in the event of a hash compromise. They also facilitate diversity and non-likability, enabling the generation of multiple distinct hashes from the same biometric input, thereby safeguarding against cross-matching across various systems. Furthermore, these methods are computationally efficient, adaptable to different biometric modalities, and resistant to collusion, rendering them suitable for robust and scalable biometric security solutions. By opting for these techniques, one can achieve a strong equilibrium of security, efficiency, adaptability, and privacy. Such advantages render them particularly appropriate for contemporary biometric systems that necessitate strong protection against emerging threats while ensuring performance and user convenience.

The experimental results

In this section, we discuss the experimental results of the proposed CanDark approach. First, we adjust the CanDark model parameters before conducting the experiments. Some of these parameters are; (i) Weight decay (set to 5 × 10− 4); the best value ranges from zero to 0.1. (ii) Momentum (set to 0.9). (iii) #epochs, we select different number of epochs (20, 30, 40, 50, and 60) for the evaluation of the CanDark model. (iv) Learning rate (LR) (set to 0.01 and 0.001).

We measure the classification performance in terms of accuracy, specificity, precision, recall, and fscore. For all metrics, a number of statistical parameters are computed such as mean, standard error of median, mean, standard deviation, mode, skewness, variance, standard error of skewness, minimum, and maximum range. The mean is the average value of the results. The standard error of mean determines the discrepancy of the results compared with the mean. The Median is the midst value of the experimental results. The mode appears most frequently in experiments. The standard deviation and variance values give an indication about the closeness of measurements around the mean value. Lower values of standard deviation and variance are preferred to higher values. Skewness measures the symmetry or asymmetry in the data distribution. The distribution is considered highly skewed if (skewness < -1 or skewness > 1), moderately skewed if (-1 < skewness < -0.5 or 0.5 < skewness < 1), and approximately symmetric if (-0.5 < skewness < 0.5). The standard error of skewness is the ratio of skewness to standard error. The range is the difference between the maximum and the minimum values. Tables 4, 5, 6, and 7 illustrate the performance of the proposed framework using Bio-Hashing, Multi-Layer Perceptron (MLP) Hashing, IoM-GRP, and IoM-URP, respectively. From these results, it could be noticed that:

Table 4.

Performance and numerical statistical analysis of the proposed framework using Bio-hashing cancelable technique.

#epochs LR Accuracy Specificity Precision Recall Fscore
20 0.01 92.44 90.91 90.80 94.05 92.40
0.001 94.19 93.18 93.2 95.24 94.12
30 0.01 92.44 97.73 97.33 86.90 91.82
0.001 97.09 98.86 98.77 95.24 96.97
40 0.01 95.24 95.45 95.24 95.7 95.24
0.001 93.60 94.32 93.98 92.86 93.41
50 0.01 91.28 87.50 87.91 95.29 91.43
0.001 96.51 96.22 96.43 96.43 96.43
60 0.01 92.44 96.13 96.10 88.10 91.93
0.001 97.67 96.59 96.51 98.81 97.65
Mean 94.2900 94.7720 94.6270 93.811 94.140
Std. Error of Mean 0.70639 1.08716 1.03297 1.1612 0.72995
Median 93.8950 96.0200 95.6700 95.240 93.765
Mode 92.44 96.59 87.91a 95.24 91.43a
Std. Deviation 2.23381 3.43790 3.26652 3.6720 2.3083
Variance 4.990 11.819 10.670 13.484 5.328
Skewness 0.321 -1.141 -1.018 -1.017 0.343
Std. Error of Skewness 0.687 0.687 0.687 0.687 0.687
Range 6.39 11.36 10.86 11.91 6.22
Minimum 91.28 87.50 87.91 86.90 91.43
Maximum 97.67 98.86 98.77 98.81 97.65
Open in a new tab

Table 5.

Performance and numerical statistical analysis of the proposed framework using MLP hashing cancelable technique.

#epochs LR Accuracy Specificity Precision Recall Fscore
20 0.01 93.6 93.11 92.8 92.82 92.86
0.001 96.3 95.19 95.37 97.6 96.5
30 0.01 93.9 93.2 92.1 92.91 92.86
0.001 96.3 95.4 95.33 97.68 96.47
40 0.01 94.77 94.32 94.12 95.24 94.67
0.001 95.93 97.62 94.25 97.62 95.91
50 0.01 93.02 97.73 97.37 88.10 92.50
0.001 93.02 93.18 92.86 92.86 92.86
60 0.01 96.51 95.45 95.35 97.62 96.47
0.001 94.77 94.32 94.12 95.24 94.67
Mean 94.7080 95.0560 94.4490 94.764 94.574
Std. Error of Mean 0.50155 0.51362 0.45703 0.99087 0.53440
Median 94.7700 94.8850 94.1850 95.240 94.670
Mode 93.02 95.45 92.86a 97.62 92.86a
Std. Deviation 1.58605 1.62421 1.44524 3.1334 1.6899
Variance 2.516 2.638 2.089 9.818 2.856
Skewness 0.005 0.663 0.690 -1.008 − 0.034
Std. Error of Skewness 0.687 0.687 0.687 0.687 0.687
Range 3.49 4.55 4.51 9.52 3.97
Minimum 93.02 93.18 92.86 88.10 92.50
Maximum 96.51 97.73 97.37 97.62 96.47
Open in a new tab

Table 6.

Performance and numerical statistical analysis of the proposed framework using IoM-GRP cancelable technique.

#epochs LR Accuracy Specificity Precision Recall Fscore
20 0.01 97.23 97.13 97.49 96.42 97.01
0.001 96.51 95.1 95.31 97.63 96.43
30 0.01 94.19 95.45 95.12 92.86 93.92
0.001 97.01 97.6 97.2 96.2 97.01
40 0.01 96.2 95.6 95.35 97.62 96.44
0.001 94.7 95.7 95.25 92.13 93.96
50 0.01 96.51 96.59 96.43 96.43 96.41
0.001 97.67 97.73 97.62 97.62 97.62
60 0.01 97.09 97.73 97.59 96.43 97.01
0.001 92.44 93.18 92.77 91.67 92.22
Mean 95.9290 96.2490 96.0530 95.597 95.820
Std. Error of Mean 0.54067 0.48146 0.50763 0.71069 0.56211
Median 96.5100 96.0200 95.8900 96.430 96.470
Mode 96.51a 95.45a 97.59 96.43 97.01
Std. Deviation 1.70974 1.52250 1.60528 2.2473 1.7775
Variance 2.923 2.318 2.577 5.051 3.160
Skewness -1.179 − 0.705 − 0.769 − 0.911 -1.173
Std. Error of Skewness 0.687 0.687 0.687 0.687 0.687
Range 5.23 4.55 4.85 5.95 5.40
Minimum 92.44 93.18 92.77 91.67 92.22
Maximum 97.67 97.73 97.62 97.62 97.62
Open in a new tab

Table 7.

Performance and numerical statistical analysis of the proposed framework using IoM-URP cancelable technique.

#epochs LR Accuracy Specificity Precision Recall Fscore
20 0.01 98.23 98.42 97.05 98.27 97.65
0.001 97.06 98.53 97.6 98.11 97.85
30 0.01 97.06 97.95 98.19 97.64 97.91
0.001 96.8 96.87 96.74 97.26 96.99
40 0.01 98.72 97.77 98.68 97.28 97.97
0.001 97.1 97.46 97.75 96.21 96.97
50 0.01 98.72 98.44 99.02 98.45 98.73
0.001 98.81 98.73 98.9 97.02 97.95
60 0.01 98.12 97.72 98.67 97.66 98.16
0.001 98.31 98.18 99.82 98.94 99.37
Mean 97.8930 98.0070 98.2420 97.684 97.955
Std. Error of Mean 0.25296 0.18006 0.30253 0.25052 0.22737
Median 98.1750 98.0650 98.4300 97.650 97.930
Mode 97.06a 96.87a 96.74a 96.21a 96.97a
Std. Deviation 0.79993 0.56941 0.95668 0.79220 0.71900
Variance 0.640 0.324 0.915 0.628 0.517
Skewness − 0.254 − 0.747 − 0.098 − 0.259 0.532
Std. Error of Skewness 0.687 0.687 0.687 0.687 0.687
Range 2.01 1.86 3.08 2.73 2.40
Minimum 96.80 96.87 96.74 96.21 96.97
Maximum 98.81 98.73 99.82 98.94 99.37
Open in a new tab
  • I.

    For Bio-Hashing method, the maximum accuracy, specificity, precision, recall, and fscore are 97.67%, 98.86%, 98.77%, 98.81%, and 97.65%, respectively. The range of accuracy, specificity, precision, recall, and fscore are 6.39, 11.36, 10.86, 11.91, and 6.22, respectively. The skewness of accuracy is (0.321) (approximately symmetric), specificity is (-1.141) (highly skewed), precision is (-1.018) (highly skewed), recall is (-1.017) (highly skewed), fscore is (0.343) (approximately symmetric). The maximum accuracy is achieved using (#epochs = 60 and LR = 0.001).

  • II.

    For MLP Hashing method, the maximum accuracy, specificity, precision, recall, and fscore are 96.51%, 97.73%, 97.37%, 97.62%, and 96.47%, respectively. The range of accuracy, specificity, precision, recall, and fscore are 3.49, 4.55, 4.51, 9.52, and 3.97, respectively. The skewness of accuracy is (0.005) (approximately symmetric), specificity is (0.663) (moderately skewed), precision is (0.69) (moderately skewed), recall is (-1.008) (highly skewed), fscore is (-0.034) (approximately symmetric). The maximum accuracy is achieved using (#epochs = 60 and LR = 0.01).

  • III.

    For IoM-GRP method, the maximum accuracy, specificity, precision, recall, and fscore are 97.67%, 97.73%, 97.62%, 97.62%, and 97.62%, respectively. The range of accuracy, specificity, precision, recall, and fscore are 5.23, 4.55, 4.85, 5.95, and 5.4, respectively. The skewness of accuracy is (-1.179) (highly skewed), specificity is (-0.705) (moderately skewed), precision is (-0.769) (moderately skewed), recall is (-0.911) (moderately skewed), fscore is (-1.173) (highly skewed). The maximum accuracy is achieved using (#epochs = 50 and LR = 0.001).

  • IV.

    For IoM-URP method, the maximum accuracy, specificity, precision, recall, and fscore are 98.81%, 98.73%, 99.82%, 98.94%, and 99.37%, respectively. The range of accuracy, specificity, precision, recall, and fscore are 2.01, 1.86, 3.08, 2.73, and 2.4, respectively. The skewness of accuracy is (-0.254) (approximately symmetric), specificity is (-0.747) (moderately skewed), precision is (-0.098) (approximately symmetric), recall is (-0.259) (approximately symmetric), fscore is (0.532) (moderately skewed). The maximum accuracy is achieved using (#epochs = 50 and LR = 0.001).

  • V.

    According to accuracy, the mean value for IoM-URP (97.89%) > IoM-GRP (95.93%) > MLP Hashing (94.7%) > Bio-Hashing (94.29%). The variance value for IoM-URP (0.64) < MLP Hashing (2.516) < IoM-GRP (2.923) < Bio-Hashing (4.99). The standard deviation value for IoM-URP (0.799) < MLP Hashing (1.586) < IoM-GRP (1.7) < Bio-Hashing (2.23). The range for Bio-Hashing (6.39) > IoM-GRP (5.23) > MLP Hashing (3.49) > IoM-URP (2.01).

  • VI.

    According to specificity, the mean value for IoM-URP (98%) > IoM-GRP (96.25%) > MLP Hashing (95.05%) > Bio-Hashing (94.77%). The variance value for IoM-URP (0.324) < IoM-GRP (2.23) < MLP Hashing (2.64) < Bio-Hashing (11.8). The standard deviation value for IoM-URP (0.569) < IoM-GRP (1.52) < MLP Hashing (1.62) < Bio-Hashing (3.44). The range for Bio-Hashing (11.36) > (IoM-GRP (4.55) and MLP Hashing (4.55)) > IoM-URP (1.86).

  • VII.

    According to precision, the mean value for IoM-URP (98.24%) > IoM-GRP (96.05%) > Bio-Hashing (94.63%) > MLP Hashing (94.45%). The variance value for IoM-URP (0.915) < MLP Hashing (2.089) < IoM-GRP (2.577) < Bio-Hashing (10.67). The standard deviation value for IoM-URP (0.9566) < MLP Hashing (1.445) < IoM-GRP (1.605) < Bio-Hashing (3.266). The range for Bio-Hashing (10.86) > IoM-GRP (4.85) > MLP Hashing (4.51) > IoM-URP (3.08).

  • VIII.

    According to recall, the mean value for IoM-URP (97.684%) > IoM-GRP (95.597%) > MLP Hashing (94.764%) > Bio-Hashing (93.811%). The variance value for IoM-URP (0.628) < IoM-GRP (5.051) < MLP Hashing (9.818) < Bio-Hashing (13.484). The standard deviation value for IoM-URP (0.7922) < IoM-GRP (2.247) < MLP Hashing (3.133) < Bio-Hashing (3.672). The range for Bio-Hashing (11.91) > MLP Hashing (9.52) > IoM-GRP (5.95) > IoM-URP (2.73).

  • IX.

    According to fscore, the mean value for IoM-URP (97.955%) > IoM-GRP (95.82%) > MLP Hashing (94.574%) > Bio-Hashing (94.14%). The variance value for IoM-URP (0.517) < MLP Hashing (2.856) < IoM-GRP (3.16) < Bio-Hashing (5.328). The standard deviation value for IoM-URP (0.719) < MLP Hashing (1.6899) < IoM-GRP (1.7775) < Bio-Hashing (2.308). The range for Bio-Hashing (6.22) > IoM-GRP (5.4) > MLP Hashing (3.97) > IoM-URP (2.4).

  • X.

    IoM-URP method is superior to all other cancelable techniques in terms of all performance metrics and statistical parameters.

In addition, Figs. 7, 8, 9, 10, 11, 12, 13) display graphical representations of the proposed framework using Bio-Hashing, MLP Hashing, IoM-GRP, and IoM-URP, respectively. Figures 8, 10, 12, and 14 depict the error bars of accuracy, specificity, precision, recall, and fscore of the proposed framework using Bio-Hashing, MLP Hashing, IoM-GRP, and IoM-URP, respectively. Error bars indicate the estimated error and the uncertainty of experiments.

Fig. 7.

Fig. 7

Open in a new tab

Graphical representation of the proposed framework using Bio-Hashing cancelable technique.

Fig. 8.

Fig. 8

Open in a new tab

Error bar of accuracy, specificity, precision, recall, and fscore of the proposed framework using Bio-Hashing cancelable technique.

Fig. 9.

Fig. 9

Open in a new tab

Graphical representation of the proposed framework using MLP Hashing cancelable technique.

Fig. 10.

Fig. 10

Open in a new tab

Error bar of accuracy, specificity, precision, recall, and fscore of the proposed framework using MLP Hashing cancelable technique.

Fig. 11.

Fig. 11

Open in a new tab

Graphical representation of the proposed framework using IoM-GRP cancelable technique.

Fig. 12.

Fig. 12

Open in a new tab

Error bar of accuracy, specificity, precision, recall, and fscore of the proposed framework using IoM-GRP cancelable technique.

Fig. 13.

Fig. 13

Open in a new tab

Graphical representation of the proposed framework using IoM-URP cancelable technique.

Fig. 14.

Fig. 14

Open in a new tab

Error bar of accuracy, specificity, precision, recall, and fscore of the proposed framework using IoM-URP cancelable technique.

The analysis of the previous measurements confirms the superiority of using IoM-URP cancelable algorithm to provide cancelability of the extracted features from the DarkNet-53 CNN model. For further analysis of the proposed framework using IoM-URP method, Figs. 15, 16, 17, 18, 19, 20, 21, 22, and 23 display the normal Q-Q plot of the accuracy, specificity, precision, recall, and fscore, respectively. The Q-Q plot determines if the measurements are normally distributed (points fall on the 45-degree line) or not. Moreover, Figs. 16, 18, 20, 22, 24 present the detrended normal Q-Q plot of the accuracy, specificity, precision, recall, and fscore, respectively. These plots are used to find the difference between the observed and the predicted measurements. The results are considered to be normally distributed if that difference lies around the zero line.

Fig. 15.

Fig. 15

Open in a new tab

Normal Q-Q plot of the accuracy of the proposed framework using IoM-URP cancelable technique.

Fig. 16.

Fig. 16

Open in a new tab

Detrended normal Q-Q plot of the accuracy of the proposed framework using IoM-URP cancelable technique.

Fig. 17.

Fig. 17

Open in a new tab

Normal Q-Q plot of the specificity of the proposed framework using IoM-URP cancelable technique.

Fig. 18.

Fig. 18

Open in a new tab

Detrended normal Q-Q plot of the specificity of the proposed framework using IoM-URP cancelable technique.

Fig. 19.

Fig. 19

Open in a new tab

Normal Q-Q plot of the precision of the proposed framework using IoM-URP cancelable technique.

Fig. 20.

Fig. 20

Open in a new tab

Detrended normal Q-Q plot of the precision of the proposed framework using IoM-URP cancelable technique.

Fig. 21.

Fig. 21

Open in a new tab

Normal Q-Q plot of the recall of the proposed framework using IoM-URP cancelable technique.

Fig. 22.

Fig. 22

Open in a new tab

Detrended normal Q-Q plot of the recall of the proposed framework using IoM-URP cancelable technique.

Fig. 23.

Fig. 23

Open in a new tab

Normal Q-Q plot of the fscore of the proposed framework using IoM-URP cancelable technique.

Fig. 24.

Fig. 24

Open in a new tab

Detrended normal Q-Q plot of the fscore of the proposed framework using IoM-URP cancelable technique.

Besides, Table 8; Fig. 25 proves the superiority of using IoM-URP over other cancelable methods through a numerical and graphical comparisons between the accuracy, specificity, precision, recall, and fscore. Finally, Table 9 provides the performance of the proposed CanDark model against state-of-the-art techniques.

Table 8.

Performance of the proposed framework under various cancelable techniques.

Cancelable technique Accuracy Specificity Precision Recall Fscore
Bio-hashing 97.67 96.59 96.51 98.81 97.65
MLP Hashing 96.51 95.45 95.35 97.62 96.47
IoM-GRP 97.67 97.73 97.62 97.62 97.62
IoM-URP 98.81 98.73 98.9 97.02 97.95
Open in a new tab

Fig. 25.

Fig. 25

Open in a new tab

Graphical comparison between accuracy, specificity, precision, recall, and fscore of the proposed framework under various cancelable techniques.

Table 9.

Performance of the proposed CanDark model against state-of-the-art techniques.

Method Accuracy Precision Recall Fscore
Irmak31 91.38 90 86 88
Akin33 98.25 100 96 98
Sitaula34 87.13 85 85 85
Altun36 96 99 97 98
Ahsan39

Batch Size = 10

Epochs = 50

LR = 0.001

 VGG16 81 86 81 80
InceptionV2 90 88 88 77
ResNet-50 56 32 56 40
ResNet-101 56 32 56 40
MobileNetV2 81 82 81 81
VGG19 93 94 94 94

Batch Size = 10

Epochs = 50

LR = 0.001

 VGG16 93 93 90 91
InceptionV2 98 98 97 98
ResNet-50 72 51 72 60
ResNet-101 72 36 50 42
MobileNetV2 99 99 99 99
VGG19 90 89 86 87
Proposed CanDark Model 98.81 98.9 97.02 97.95
Open in a new tab

The proposed method outperforms all other methods across most criteria, with the exception of Ahsan39 (using MobileNetV2). This discrepancy arises because Ahsan39 employs an 80% training rate, while our proposed method uses a 70% training rate. However, we also tested our CanDark model using an 80% training rate, and the results were as follows: accuracy = 99.38%, Precision = 99.53%, recall = 99.16% and fscore = 99.34%.

Conclusion

As COVID-19 virus infections drop globally, the Monkeypox virus is steadily emerging. Early detection of them may be achievable with the help of AI-based detection approaches based on deep learning. This study used 770 skin pictures, which were split into four categories: measles, chickenpox, Monkeypox, and normal cases. The proposed CanDark model is a complete paradigm that uses cancellable biometrics to identify cases of Monkeypox. It relies on using the DarkNet-53 CNN model to extract features from images of Monkeypox skin. Then, cancellable biometric techniques including bio-hashing, MLP, IoM-GRP, and IoM-URP hashing techniques can be used to protect the output information. The proposed framework achieved an accuracy of 98.81%, a specificity of 98.73%, a precision of 98.9%, a recall of 97.02%, and fscore of 97.95%. We can enlarge the data set and evaluate the suggested model’s performance in subsequent work. Moreover, we can test the effectiveness of Monkeypox case detection and classification using various pre-trained models and a scratch CNN models.

Author contributions

All authors of this research paper have directly participated in the planning, execution, or analysis of this study; All authors of this paper have read and approved the final version submitted;

Funding

no funding was received for conducting this study.

Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooperation with The Egyptian Knowledge Bank (EKB).

Data availability

Data set available on: https://www.kaggle.com/datasets/dipuiucse/monkeypoxskinimagedataset, and used by (bala et al., 2023monkeynet) MonkeyNet: A robust deep convolutional neural network for monkeypox disease detection and classification. 10.1016/j.neunet.2023.02.022.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent

Informed consent was obtained from all individual participants included in the study.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

  • 1.Antunes, F., Cordeiro, R. & Virgolino, A. Monkeypox: From a neglected tropical disease to a public health threat. Infect. Dis. Rep.14(5), 772–783 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Control P. ASSESSMENT and Monkeypox multi-country outbreak (2022).
  • 3.Nakazawa, Y. et al. Phylogenetic and ecologic perspectives of a monkeypox outbreak, southern Sudan, 2005. Emerg Infect Dis.19(2) (2013). [DOI] [PMC free article] [PubMed]
  • 4.Ladnyj, I., Ziegler, P. & Kima, E. A human infection caused by monkeypox virus in Basankusu Territory, Democratic Republic of the Congo. Bull. World Health Organ.46(5) (1972). [PMC free article] [PubMed]
  • 5.Marennikova, S., Šeluhina, E., Ceva, N., Čimiškjan, K. & Macevič, G. Isolation and properties of the causal agent of a new variola-like disease (monkeypox) in man. Bull. World Health Organ.46(5) (1972). [PMC free article] [PubMed]
  • 6.Yinka-Ogunleye, A. et al. Outbreak of human monkeypox in Nigeria in 2017–18: A clinical and epidemiological report. Emerg. Infect. Dis.19(8), 872–879 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Lourie, B. et al. Human infection with monkeypox virus: Laboratory investigation of six cases in West Africa. Emerg. Infect. Dis.46(5) (1972). [PMC free article] [PubMed]
  • 8.Tiee, M., Harrigan, R., Thomassen, H. & Smith, T. Ghosts of infections past: Using archival samples to understand a century of monkeypox virus prevalence among host communities across space and time. Bull. World Health Organ.5(1) (2018). [DOI] [PMC free article] [PubMed]
  • 9.Magnus, P., Andersen, E., Petersen, K. & Birch-Andersen, A. A pox‐like disease in cynomolgus monkeys. Acta Pathol. Microbiol. Scand.46(2), 156–176 (1959). [Google Scholar]
  • 10.Parker, S. & Buller, R. A review of experimental and natural infections of animals with monkeypox virus between 1958 and 2012. Future Virol.8(2), 129–157 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Simpson, K. et al. Human monkeypox–after 40 years, an unintended consequence of smallpox eradication. Vaccine38(33), 5077–5081 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Fenner, F. and p. sciences, Smallpox: Emergence, global spread, and eradication. Hist. Philos. Life Sci. 397–420 (1993). [PubMed]
  • 13.Alakunle, E., Moens, U., Nchinda, G. & Okeke, M. Monkeypox virus in Nigeria: Infection biology, epidemiology, and evolution. Viruses12(11) (2020). [DOI] [PMC free article] [PubMed]
  • 14.Vaughan, A. et al. Two cases of monkeypox imported to the United Kingdom, September 2018. Euro Surveill23(38) (2018). [DOI] [PMC free article] [PubMed]
  • 15.Yong, S. et al. Imported Monkeypox, Singapore. Travel Med. Infect. Dis.26, 8 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Costello, V. et al. Imported monkeypox from international traveler, Maryland, USA, 2021. Emerg. Infect. Dis.28, 5 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Guarner, J. et al. Monkeypox transmission and pathogenesis in prairie dogs. Emerg. Infect. Dis.10, 3 (2004). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Russo, A. T. et al. An overview of tecovirimat for smallpox treatment and expanded anti-orthopoxvirus applications. CDC19(3), 331–344 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.https://www.cdc.gov/poxvirus/mpox/response/2022/world-map.html, october (2023).
  • 20.Reynolds, M. G. et al. Spectrum of infection and risk factors for human monkeypox, United States. Emerg. Infect. Dis.13, 9 (2007). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Saad, W., Shalaby, W., Shokair, M., Sayed, F. & Abdellatef, E. COVID-19 classification using deep feature concatenation technique. J. Ambient Intell. Hum. Comput.13, 2025–2043 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Ravì, D. et al. Deep learning for health informatics. IEEE J. Biomed. Health Inf.21(1), 4–21 (2016). [DOI] [PubMed] [Google Scholar]
  • 23.Ciregan, D., Meier, U. & Schmidhuber, J. Multi-column deep neural networks for image classification. In IEEE Conference on Computer Vision and Pattern Recognition, 3642–3649 (2012).
  • 24.Özkaya, U., Öztürk, Ş., Budak, S., Melgani, F. & Polat, K. Classification of COVID-19 in chest CT images using convolutional support vector machines. Electr. Eng. Syst. Sci. (2020).
  • 25.Cohen, J. P. et al. Predicting covid-19 pneumonia severity on chest x-ray with deep learning. Cureus12(7) (2020). [DOI] [PMC free article] [PubMed]
  • 26.Khan, G. & Ullah, I. Efficient technique for monkeypox skin disease classification with clinical data using pre-trained models. J. Innov. Image Process.5(2), 192–213 (2023). [Google Scholar]
  • 27.Özkan, İ. & M. ÇELİK and Detection of monkeypox among different pox diseases with different pre-trained deep learning models. J. Inst. Sci. Technol.13(1), 10–21 (2023). [Google Scholar]
  • 28.Pal, M. et al. Deep and transfer learning approaches for automated early detection of monkeypox (mpox) alongside other similar skin lesions and their classification. ACS Omega23(35), 31747–31757 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Pal, M. et al. COVID-19 prognosis from chest X-ray images by using deep learning approaches: A next generation diagnostic tool. J. Pure Appl. Microbiol.17(2), 919–930 (2023). [Google Scholar]
  • 30.Sahin, V., Oztel, I. & Oztel, G. Human monkeypox classification from skin lesion images with deep pre-trained network using mobile application. J. Med. Syst.46, 11 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Irmak, M., Aydin, T. & Yağanoğlu, M. Monkeypox skin lesion detection with MobileNetV2 and VGGNet models. In Medical Technologies Congress (TIPTEKNO), 1–4 (2022).
  • 32.Alcalá-Rmz, V., Villagrana-Bañuelos, K., Celaya-Padilla, J., Galván-Tejada, J. & Gamboa-Rosales, H. & Galván-Tejada, C. Convolutional neural network for monkeypox detection. In International Conference on Ubiquitous Computing and Ambient Intelligence, 89–100 (2022).
  • 33.Akin, K., Gurkan, C., Budak, A. & Karataş, H. Classification of monkeypox skin lesion using the explainable artificial intelligence assisted convolutional neural networks. Eur. J. Sci. Technol.40, 106–110 (2022). [Google Scholar]
  • 34.Sitaula, C. & Shahi, T. Monkeypox virus detection using pre-trained deep learning-based approaches. J. Med. Syst.46, 11 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Ali, S. N. et al. Monkeypox skin lesion detection using deep learning models: A feasibility study. Comput. Vis. Pattern Recognit. (2022).
  • 36.Altun, M. et al. Monkeypox detection using CNN with transfer learning. Sensors23, 4 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Islam, T., Hussain, M., Chowdhury, F. & Islam, B. Can artificial intelligence detect monkeypox from digital skin images? Bioexiv (2022).
  • 38.Haque, M., Ahmed, M., Nila, R. & Islam, S. Classification of human monkeypox disease using deep learning models and attention mechanisms. Image Video Process. (2022).
  • 39.Ahsan, M. M. et al. Transfer learning and local interpretable model agnostic based visual approach in Monkeypox Disease Detection and classification: A deep learning insights. Image Video Process. (2022).
  • 40.https://www.kaggle.com/datasets/dipuiucse/monkeypoxskinimagedataset.
  • 41.Rathgeb, C., Breitinger, F. & Busch, C. Alignment-free cancelable iris biometric templates based on adaptive bloom filters. In International Conference on Biometrics (ICB), 1–8 (2013).
  • 42.Manisha & Kumar, N. Cancelable biometrics: A comprehensive survey. Artif. Intell. Rev.53, 3403–3446 (2020). [Google Scholar]
  • 43.Shahreza, H., Hahn, V. & Marcel, S. Mlp-hash: Protecting face templates via hashing of randomized multi-layer perceptron. Comput. Vis. Pattern Recognit. (2022).
  • 44.Tan, F., Ong, T., Tee, C. & Teoh, A. Image hashing enabled technique for biometric template protection. TENCON IEEE Region 10 Conference, 1–5 (2009).
  • 45.Yee, L. & De Silva, L. Application of multilayer perceptron network as a one-way hash function. In Proceedings of the 2002 International Joint Conference on Neural Networks. IJCNN’02, Vol. 2, No. 2, 1459–1462 (2002).
  • 46.Jin, Z., Hwang, J., Lai, Y., Kim, S. & Teoh, A. Ranking-based locality sensitive hashing-enabled cancelable biometrics: Index-of-max hashing. IEEE Trans. Inform. Forensics Secur.. 13(2), 393–407 (2017). [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

Data set available on: https://www.kaggle.com/datasets/dipuiucse/monkeypoxskinimagedataset, and used by (bala et al., 2023monkeynet) MonkeyNet: A robust deep convolutional neural network for monkeypox disease detection and classification. 10.1016/j.neunet.2023.02.022.

Articles from Scientific Reports are provided here courtesy of Nature Publishing Group

RESOURCES