A novel dataset and local interpretable model-agnostic explanations (LIME) for monkeypox prediction

Abstract

In image-based diagnosis, machine learning recently showed great potential, particularly in the detection of cancer, the identification of tumour cells, and the diagnosis of COVID-19. Similar methods could be used to detect monkeypox on human skin, however there isn’t a public dataset with data on monkeypox that can be used to train and evaluate machine learning models. In order to address this, the dataset “Monkeypox2022” has been created and made available on GitHub (https://github.com/Angmo21720/data-set). Images from internet and open-source sources were used to construct the collection; these sources did not impose any restrictions on use, including commercial use. Additionally, the enhanced VGG16 model was proposed and ran two Studies for comparison. According to the findings, the model’s AUC Score for identifying monkeypox patients in Study One was 0.972, while its AUC Score for identifying such patients in Study Two was 0.748. Deeper understanding of the characteristics that distinguish a monkeypox infection is provided by using Local Interpretable Model-Agnostic Explanations (LIME), which is used to decipher predictions and reveal the features utilized by a model in decision-making.

Keywords

LIME monkeypox virus machine learning monkey pox data set transfer learning

1. Introduction

The COVID-19 pandemic in 2020 was a global alert and a call to action, but the emergence of monkey pox in 2022 in several countries highlighted another global threat [1]. Monkeypox is a viral disease caused by the Zoonotic Orthopox virus, belonging to the Poxviridae family and Orthopox virus genus, closely related to cowpox and smallpox. The virus primarily spreads from monkeys and rodents to humans, but transmission between humans is also common [2]. Monkey pox was first discovered in a monkey in a laboratory in Copenhagen, Denmark in 1958 [3]. The preliminary human case of monkeypox was reported in 1970 [4] and it was speculated that the people living near tropical rainforests are at risk of contracting monkey pox through close contact with infected individuals, animals, or objects [5]. Symptoms of monkey pox include fever, body pain, fatigue, and later, the appearance of a red bump on the skin [6].

Despite being less contagious than COVID-19, the incidence of monkey pox is on the rise. In the 1990s, there were only a few dozen cases reported annually in West and Central Africa [7]. However, by 2020, the number of cases had risen to 5,000. Previously thought to only occur in Africa, cases of monkey pox have now been reported in other part of world, causing growing fear and concern among the public, as seen through social media posts [8]. No specific treatment has been recommended for monkey pox, but two oral medications, Brincidofovir and Tecovirimat, are recommended for its treatment and prevention [9, 10]. In other countries, smallpox vaccines are given to those with monkey pox [11]. Diagnosing monkey pox requires examination of skin lesions and a patient’s exposure history, with electron microscopy being the definitive method. Polymerase Chain Reaction (PCR) [12] commonly used in COVID-19 diagnosis may also be used to confirm the presence of the virus [13].

No research has been found that shows the potential of Machine Learning (ML) in detecting monkeypox through image processing. The lack of research on using ML for diagnosing monkeypox through image processing is due to the fact that no public data set for ML model training and evaluation is available.Widespread virus exposure in many countries could lead to development of an appropriate ML method with a data set in the future.

There is limited research on using ML for diagnosing monkeypox through image processing due to the lack of a public dataset. Taking all these possibilities into consideration, we concluded that there is an imperative need to compile a data set that contains images of patients who are suffering from the monkeypox disease. Dr. Joseph Cohen’s creation of the Monkey pox data set at the beginning of the COVID-19 outbreak served as an inspiration for us when we decided to create the monkeypox data set. Dr. Cohen assembled the data set of 98 chest X-ray samples from a variety of different sources, including websites and journals [23]. Taking this into consideration, we have compiled a dataset of 164 images of patients suffering from monkeypox, which after data augmentation techniques, resulted in 553 total samples. This difference is due to the fact that our data set was created more recently. Different studies carried out in the beginning of COVID-19 make use of restricted datasets and emphasize the importance of transfer learning approaches [14, 15, 16, 17, 18]. These studies were undertaken by numerous researchers. Our dataset is expected to aid researchers and practitioners in constructing a model for diagnosing monkeypox. The technical contribution consists:

i)
The first publicly accessible collection of images of monkeypox was made by combining images acquired from many sources (including websites and news organisations). It is stored in the GitHub repository: https://github.com/Angmo21720/data-set.
ii)
To detect patients with monkeypox based on photos, a low-modified VGG16 model was used.
iii)
Locally Interpretable Model-agnostic Explanations (LIME) are used to create the explanation of analysis in order to validate the findings.

The manuscript is organized into six sections, including a review of deep learning, transfer learning, and LIME in Section 2, a condensed description of the experiment’s methodology in Section 3, findings in Section 4, discussion in Section 5, and a summary of results and potential avenues for further investigation in Section 6.
2. Related work

Machine Learning (ML) has proven useful in diverse fields, including medical imaging and disease diagnosis. ML’s features allow it to effectively and efficiently solve medical imaging problems. As an example, Miranda and Felipe (2015) developed a computer-aided diagnostic system for breast cancer diagnosis utilizing fuzzy logic, which mimics the thinking of radiologists and saves time. The system outputs cancer detection results based on specific criteria such as contour, shape and density [18].

Ardakani et al. (2020) assessed 10 distinct ML models using a small data set consisting of 108 COVID-19 samples and 86 Non-COVID-19 samples, and attained an accuracy of 99 percent [19]. Further, Wang et al. (2020) achieved 73.1 percent accuracy based on a proposed inception based novel model [20]. Sandeep et al. (2022) used CNN model for detecting a number of skin ailments and showed that using an existing VGG Net version, it is possible to diagnose skin diseases with a degree of accuracy of 71%. In contrast, they recommended strategy that produces the best results, displaying higher performance by reaching accuracy of 78%. Velasco et al. proposed an app and discovered an accuracy of about 94.4 percent in identifying people with chickenpox symptoms [21]. Roy et al. (2019) were able to recognize a number of skin diseases, including acne, chickenpox, cellulitis, candidiasis, and others, using a variety of segmentation algorithms [22].

The rapid spread of the monkeypox virus across multiple countries highlights the importance of detecting individuals with potential infection. The increasing pressure on clinical diagnostics due to the epidemic has led medical experts to believe that AI technology could help ease this burden by analysing visual data [13]. The aim is to enhance the level of medical care provided to COVID-19 patients in hospitals through these measures. However, at present, there is no publicly available data set for monkeypox. The potential advantages of using an AI-based method to rapidly detect and eradicate the monkeypox illness are therefore difficult to understand [14, 15, 16, 17, 18].

A crucial component of the research gap that is being discussed here was collecting patient photographs that show monkeypox. However, lesser availability of the photographs of patients having monkeypox could be an issue but for initial test cases, this should not pose any issue. At the same time, the content of the database will be continuously updated with new data that has been given by numerous organisations based in numerous nations all over the world.

3. Methodology

This section covers the various steps followed in methodology of collection of the dataset and application of models.

3.1 Data collection

Figure 1.

Various steps followed for data collection.

Following processes are followed in the process of data collection.

The dataset was obtained from multiple online sources due to the absence of a trustworthy medical dataset available to the public. The initial dataset was constructed through the use of Google search engine as illustrated in Fig. 1, which outlines the different steps involved in the data search process. A more in-depth explanation of these steps follows.

ii)

The process of gathering data for non-monkeypox samples uses a comparable method and involves searching for terms such as “Blisters” and “Macules”, as well as normal images such as those of faces, legs and hands without any signs of illness.

iii)

Dataset size is increased by manually collecting more normal photos from individuals without skin disease symptoms, who have given their consent through a signed permission form.

Figure 2.

Image sample set [15, 29].

Figure 2 demonstrates the dataset of the images collected from various sources.

Table 1 displays the features of the datasets generated in this study. However, there are only 21720 samples., classical machine learning and transfer learning can be used to create a disease diagnosis model, as demonstrated by earlier COVID-19 studies that used as few as 40–100 samples and deep learning models [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24].

Table 1

Total samples

Data set	Total sample
Skin lesion blisters (SLB)	74
Skin lesion macules (SLM)	88
Skin lesion papules (SLP)	35
Monkeypox	164
SLB augmented	121
SLM augmented	113
SLP augmented	62
Monkeypox augmented	553
E samples	1210

3.2 Data augmentation

For effective training, deep learning networks like CNN, RNN, and GANs require vast data sets. However, transfer learning and conventional ML can produce good results with fewer samples. For data augmentation in this study, the Keras Image Data Generator is used. It provides a variety of options, including rotation, scaling, and flipping [25], to improve image data. Following advice from the reference [26], Table 2 displays the parameters that were employed.

Table 2
Percentage of applied data augmentation

Technique	Percentage
Geometric transformation	Up to 5%
Color space transformation	Up to 45%
Kernel filters	2%
Mixing images	5%

3.3 Transfer learning approaches

Figure 3.

Proposed modified VGG16 model.

In the beginning, transfer learning was used in a pilot test to gauge how well machine learning models performed on the generated data set. Initial testing was conducted with a customized VGG16 model [27]. A pre-trained architecture, an updated layer, and a prediction class make up the core model. High-dimensional features are recognized by the pre-trained architecture, and the updated layer then processes the fresh data. The used modified VGG16 model is depicted in Fig. 3. It comprises 16 CNN layers with unique filter widths and stride values [28]. The model starts with an input layer, followed by two 3 $\times$ 3 filter convolutional layers, a Max Pooling layer, and this pattern repeats until the modified layer section. The architecture is flattened by the modified layer, followed by three dense layers and a dropout layer [29]. A well-known model-independent technique for analyzing the projections of machine learning models is SHAP (Shapley Additive exPlanations). It is built on the concept of cooperative games and gives each input feature a unique Shapley coefficient that measures the influence of each input characteristic on the model’s estimation. These Shapley values can be combined to understand how each factor affects the model’s output. In order to analyze complicated black-box models and to help with decision-making, SHAP offers a strategy that is logical and consistent.

A SHAP extension made exclusively for deep learning models is called DeepSHAP. It makes use of a SHAP algorithm variation that takes into account the deep neural network input spaces’ multidimensionality and frequently complex structures. DeepSHAP makes it easier to assign estimates to multiple input characteristics, allowing researchers and practitioners to understand the importance of various features in the context of deep learning models. This transparency makes it easier to troubleshoot and improve these models and fosters trust in their practical use.

DeepLIFT is a deep learning interpretation method that tries to explain specific predictions by linking the model’s output to specific input features. This is done by comparing the forward pass activations of each neuron or network node to reference activations produced from a baseline input. DeepLIFT can emphasize the most important input data elements that significantly influence the prediction, making it simpler to understand how the model decides what to predict. This approach is very useful for understanding the importance of features in intricate deep learning models.

Another method for model-independent interpretability that sheds light on machine learning model predictions is CXplain. To give a comprehensive comprehension of each forecast, it combines Shapley values and contrastive explanations. CXplain pinpoints the fundamental traits that result in distinctive predictions by contrasting a specific instance to a comparable alternative. This technique helps in understanding the rationale behind particular choices and identifying potential biases or inaccuracies in the model’s predictions. CXplain is a useful tool for assuring the equity and openness of AI applications.

In the first phase of parameter tuning, the batch size, number of epochs, and learning rate are adjusted to optimize the performance of the proposed model. The initial choices for the study were made based on the experimental parameters given in Table 3.

Table 3

Parameters for the VGG16 model

Parameters	Batch size	Learning rate	Number of epochs
Study 1	[5, 10, 15, 20]	[0.1, 0.01, 0.001]	[30, 35, 40, 45]
Optimal Parameters for Study 1	10	0.001	35
Study 2	[30, 50, 70, 100]	[0.1, 0.01, 0.001]	[50, 70, 100, 150]
Optimal Parameters for Study 2	30	0.001	100

The Adam optimization algorithm, known for its superior performance compared to other algorithms, was used to minimize the model loss. Adam often shows remarkable results in binary image classification, making it a useful choice for this study [30].

3.4 LIME as explainable AI

LIME (Local interpretable model agonistic explanations) is an effective technique that aids in the analysis of model predictions and offers perceptions into the “Blackbox” of CNN models [31]. Its effectiveness in describing image classification complexities has led to widespread use in recent years [32]. LIME employs super pixels, segments of an image with a high amount of data, to assist in initial predictions. Table 4 shows the LIME parameters used in this research to compute super pixel values. These parameters have been proven beneficial in various image prediction studies [33].

Table 4
Parameter used to identify super pixels

Function	Value
Maximum distance	150
Kernel size	4
Ratio	0.2

4. Results and discussion

4.1 Experiment setup

In order to conduct the experiment, a standard laptop with Windows 10, 16 GB of RAM, and an Intel Core I7 processor was used. The entire experiment was conducted five times, and the outcome that is being presented here is the average of the outcomes from all five computing runs.

4.2 Dataset

Table 5
Samples of images for dataset

Study	Label	Train set	Test set	Total
One	Monkeypox	132	32	164
	Skin lesion	157	40	197
	Total	289	72	361
Two	Monkeypox augmented	442	111	553
	Skin lesions augmented	236	60	296
	Total	678	171	849

The data set used for the investigation is shown in Table 5. It includes 43 samples of monkeypox and 47 samples of chickenpox for Study One and 587 samples of augmented monkeypox and 1167 additional samples for Study Two. The data samples for Chickenpox, Measles, and Normal were combined to form the “Other” class. Training, testing ratio is 80:20, a common procedure in machine learning [34, 35, 36].

4.3 Performance evaluation

Key performance indicators used to assess the efficacy of a machine learning model include accuracy, recall, precision, F1-score, sensitivity, and specificity [37, 38, 39, 40, 41, 42].

The proportion of accurate predictions the model makes compared to all of its forecasts is known as accuracy. Recall quantifies the percentage of true positive predictions to all real positive cases, while precision determines the ratio of true positive forecasts to all positive predictions generated by the model. The F1-score is a composite metric that combines precision and recall harmonic means to get a single number that sums up the model’s overall performance. Sensitivity measures the proportion of true positive forecasts to all real positive situations, whereas specificity gauges the proportion of real negative predictions to all real negative cases. These measurements assist in offering a thorough picture of a model’s effectiveness and data classification accuracy.

Table 6 showcases the performance metrics (accuracy, recall, precision, F1 score, sensitivity, and specificity) for Study One and Study Two with 95% confidence interval (CI) due to limited samples. Study One outperforms Study Two in terms of accuracy with up to 9% higher accuracy in the train set. The models exhibit high sensitivity but low specificity. Despite this, their overall performance is still noteworthy.

Table 6
Results of performance metrics of studies one and two

Study	Accuracy (%)	Precision	Recall	F1-score	Sensitivity	Specificity
Study 1 train set	0.97 $\pm$ 0.018	0.97 $\pm$ 0.018	0.97 $\pm$ 0.018	0.97 $\pm$ 0.018	0.973 $\pm$ 0.017	0.97 $\pm$ 0.018
Study 1 test set	0.83 $\pm$ 0.085	0.88 $\pm$ 0.072	0.83 $\pm$ 0.085	0.83 $\pm$ 0.851	0.66 $\pm$ 0.12	0.89 $\pm$ 0.008
Study 2 train set	0.88 $\pm$ 0.008	0.86 $\pm$ 0.009	0.87 $\pm$ 0.008	0.86 $\pm$ 0.009	0.83 $\pm$ 0.010	0.89 $\pm$ 0.008
Study 2 test set	0.78 $\pm$ 0.022	0.75 $\pm$ 0.023	0.75 $\pm$ 0.023	0.75 $\pm$ 0.023	0.650 $\pm$ 0.28	0.83 $\pm$ 0.019

Figure 4.

Accuracy and loss curves for study one and two.

Figure 5.

Visualization of confusion matrices.

Figure 4 depicts the degree of accuracy and loss that the modified VGG16 model had throughout each epoch when it was applied to (a) Study One and (b) Study Two. Additionally, the results indicate that the model in Study One performed better than the model in Study Two as the accuracy for Study One was higher and reached its maximum earlier. Overfitting can also be observed in Study Two as the accuracy on the validation set decreases after reaching its maximum value at 100 epochs. These results provide a visual representation of the model’s performance and help to highlight the strengths and weaknesses of each model.

The results of the performance evaluation of the proposed model are depicted in Fig. 5, which shows the confusion matrices of both Study One and Study Two. Study One has the lowest error rate (2.7%) among all the studies. On the other hand, Study Two has the highest error rate (12.33%) due to the unbalanced ratio of the data set. The monkeypox cases to other cases ratio was 1:1.98, which could have contributed to the higher rate of misclassifications in Study Two. The figures provide a visual representation of the model’s accuracy and helps to identify the areas for improvement.

Figure 6.

AUC curves for (a) study one, and (b) study two.

Figure 7.

Classification results.

The modified VGG16 models’ ROC (Receiver Operating Characteristics) curve analysis results are shown in Fig. 6. It includes the AUC, TPR, and FPR metrics (True Positive Rate, False Positive Rate). Study One’s train set had the highest AUC at 0.972, while Study Two’s test set had the lowest at 0.748. A high AUC (close to 1) indicates better model performance, while a low AUC (close to 0.5) shows poor performance. The low AUC score of 0.748 in Study Two’s test set shows that the model’s prediction capabilities are unreliable.

Figure 7 shows the classification results found by the proposed models using LIME, depicted visually.

In this research, we created a new dataset for categorizing Monkeypox infections using image analysis techniques. A modified version of VGG16 was used and its accuracy was tested in two separate studies, resulting in an accuracy of roughly 0.83 to 0.085 on a small dataset and 0.78 with 0.022 on an unbalanced dataset. The model’s reliability was also analyzed using LIME, which is a popular explainable AI approach. The new dataset provides a valuable opportunity for further research and development of image-based tools for Monkeypox diagnosis. However, there is currently no accessible image dataset for monkeypox to compare the model’s performance with others. Following are few of the drawbacks of the study.

Lack of large and diverse data sets: The modest size of the data sets used in the research can influence the model’s generalization capacity, which could cause it to perform poorly on larger and more varied data sets.

ii)

Limited generalizability: Since the study only considered the classification of monkeypox, other types of skin illnesses may not be included in the model’s scope of applicability.

iii)

Inadequate validation: The model’s validation was only conducted using the test set, which had a small sample size, and as a result, the outcomes might not accurately reflect the model’s genuine performance.

iv)

Explainability limitations: Although LIME was used to explain the model’s predictions, its interpretation might still be arbitrary and might not fully explain the model’s behavior.

As a result, additional study is required to overcome these constraints, which can be accomplished by gathering more varied data sets, verifying the model using larger and more varied data sets, and developing more sophisticated interpretability approaches.

5. Conclusion and future scope

The goal of the research was to address the shortage of data on patient photos infected with the monkeypox virus by creating a new, publicly available dataset obtained from open sources. The modified VGG16 model was used in both studies, and results showed the ability of the model to differentiate between patients with monkeypox symptoms and those without with accuracy ranging from 78% to 97%. An explainable AI approach LIME was used to provide insights into the model’s predictions. The finding inspires future research to employ transfer learning in clinical diagnostics and demonstrates the promise of AI-based technology in the early detection and prevention of monkeypox. The strategy described in the text for the scant data and methods of analysis can be helpful to medical professionals. To further improve the model, continuous collection of new patient data, evaluation on imbalanced data, comparison with other studies, and development of a mobile-based tool are recommended.

Ethical approval

This article does not contain any studies with animals performed by any of the authors.

Funding

This research is not funded by a government or non-governmental organization.

Data availability

Data is available as dataset “Monkeypox2022” and is made available on GitHub (https://github.com/ Angmo21720/data-set).

Author contributions

In this research all the authors has equal contribution.

Footnotes

Conflict of interest

The author declare that they have no conflict of interest.

References

McCollum

Damon

. Human monkeypox. Clinical infectious diseases. 2014; 58(2): 260-267.

Alakunle

Moens

Nchinda

Okeke

. Monkeypox virus in Nigeria: infection biology, epidemiology, and evolution. Viruses. 2020; 12(11): 1257.

Moore

Rathish

Zahra

. Mpox (Monkeypox). In: StatPearls. Treasure Island (FL): StatPearls Publishing; 2023 Jan. PMID: 34662033. Available from: https://www.ncbi.nlm.nih.gov/books/NBK574519/.

Nolen

Osadebe

Katomba

Likofata

Mukadi

Monroe

Doty

Hughes

Kabamba

Malekani

Bomponda

. Extended human-to-human transmission during a monkeypox outbreak in the Democratic Republic of the Congo. Emerging Infectious Diseases. 2016 Jun; 22(6): 1014.

Nguyen

Ajisegiri

Costantino

Chughtai

MacIntyre

. Reemergence of human monkeypox and declining population immunity in the context of urbanization, Nigeria, 2017–2020. Emerging Infectious Diseases. 2021 Apr; 27(4): 1007.

Mpox signs and symptoms. Available from: https://www.cdc.gov/poxvirus/mpox/symptoms/index.html.

Doucleff

. Scientists warned us about monkeypox in 1988. Here’s why they were right. Available from: https://www.npr.org/sections/goatsandsoda/2022/05/27/1101751627/.

Multi-country monkeypox outbreak in non-endemic countries. 2022; Available from: https://www.who.int/emergencies/disease-outbreak-news/item/2022-DON385.

Rizk

Lippi

Henry

Forthal

Rizk

. Prevention and treatment of monkeypox. Drugs. 2022 Jun; 82(9): 957-63.

10.

Adler

Gould

Hine

Snell

Wong

Houlihan

Osborne

Rampling

Beadsworth

Duncan

Dunning

. Clinical features and management of human monkeypox: a retrospective observational study in the UK. The Lancet Infectious Diseases. 2022 Aug 1; 22(8): 1153-62.

11.

Park

. There’s already a monkeypox vaccine. But not everyone may need it. 2022; Available from: https://time.com/6179429/monkeypox-vaccine/.

12.

Signs and Symptoms. Available from: https://www.nj.gov/health/monkeypox/about/symptoms/index.shtml.

13.

Ahsan

Gupta

Islam

Sen

Rahman

Shakhawat Hossain

. COVID-19 symptoms detection based on nasnetmobile with explainable ai using various imaging modalities. Machine Learning and Knowledge Extraction. 2020 Oct 29; 2(4): 490-504.

14.

Ahsan

Alam

Trafalis

Huebner

. Deep MLP-CNN Model Using Mixed-Data to Distinguish between COVID-19 and Non-COVID-19 Patients. Symmetry. 2020; 12(9): 1526.

15.

Ahsan

Ahad

Soma

Paul

Chowdhury

Luna

Yazdan

Rahman

Siddique

Huebner

. Detecting SARS-CoV-2 from chest X-Ray using artificial intelligence. IEEE Access. 2021 Feb 23; 9: 35501-13.

16.

Ahsan

Nazim

Siddique

Huebner

. Detection of COVID-19 Patients from CT Scan and Chest X-ray Data Using Modified MobileNetV2 and LIME. Healthcare. 2021; 9(9): 1099. doi: 10.3390/healthcare9091099.

17.

Ahsan

Nazim

Siddique

Huebner

. Detection of COVID-19 Patients from CT Scan and Chest X-ray Data Using Modified MobileNetV2 and LIME. Healthcare. 2021; 9(9): 1099. doi: 10.3390/healthcare9091099.

18.

Miranda

Felipe

. Computer-aided diagnosis system based on fuzzy logic for breast cancer categorization. Computers in Biology and Medicine. 2015 Sep 1; 64: 334-46.

19.

Ardakani

Kanafi

Acharya

Khadem

Mohammadi

. Application of deep learning technique to manage COVID-19 in routine clinical practice using CT images: Results of 10 convolutional neural networks. Computers in Biology and Medicine. 2020 Jun 1; 121: 103795.

20.

Wang

Lin

Wong

. COVID-net: A tailored deep convolutional neural network design for detection of COVID-19 cases from chest X-ray images. Scientific Reports. 2020 Nov 11; 10(1): 19549.

21.

Sandeep

Vishal

Shamanth

Chethan

. Diagnosis of Visible Diseases Using CNNs. In: Goyal

Gupta

Mirjalili

Trivedi

, editors. Proceedings of International Conference on Communication and Artificial Intelligence. Lecture Notes in Networks and Systems, vol. 435. Singapore: Springer; 2022. doi: 10.1007/978-981-19-0976-4_38.

22.

Roy

Chaudhuri

Ghosh

Dutta

Chakraborty

Sarkar

. Skin Disease detection based on different Segmentation Techniques. In: 2019 International Conference on Opto-Electronics and Applied Optics (Optronix). IEEE; 2019 Mar 18. pp. 1-5.

23.

Cohen

Morrison

Dao

. COVID-19 Image Data Collection. 2020; arXiv preprint arXiv:2003.11597.

24.

Narin

Kaya

Pamuk

. Automatic detection of corona virus disease (covid-19) using x-ray images and deep convolution neural networks. Pattern Analysis and Applications. 2021 24(3): 1207-1220.

25.

Image Data Generator. 2022; Available from: https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image/ImageDataGenerator.

26.

Bhattiprolu

. Data augmentation. 2020; Available from https://github.com/bnsreenu.

27.

Qassim

Verma

Feinzimer

. Compressed residual-VGG16 CNN model for big data places image recognition. In: 2018 IEEE 8th Annual Computing and Communication Workshop and Conference (CCWC). Las Vegas, NV, USA: 2018. pp. 169-175. doi: 10.1109/CCWC.2018.8301729.

28.

Ahmed

Karim

AAA

. The Impact of Filter Size and Number of Filters on Classification Accuracy in CNN. In: 2020 International Conference on Computer Science and Software Engineering (CSASE). Duhok, Iraq: 2020. pp. 88-93. doi: 10.1109/CSASE48920.2020.9142089.

29.

Bergstra

Bengio

. Random Search for Hyper-Parameter Optimization. Journal of Machine Learning Research. 2012; 13(2).

30.

Perez

Wang

. The effectiveness of data augmentation in image classification using deep learning. 2017; arXiv preprint arXiv:1712.04621.

31.

Ribeiro

Singh

Guestrin

. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016. pp. 1135-1144.

32.

Cian

Gemert

Lengyel

. Evaluating the performance of the lime and grad-cam explanation methods on a Lego multi-label image classification task. 2020. arXiv preprint arXiv:2008.01584.

33.

Pan

Xiao

Han

Zhang

Jiang

Chen

Zhou

Bao

Xie

. Prognostic Assessment of COVID-19 in the Intensive Care Unit by Machine Learning Methods: Model Development and Validation. J Med Internet Res. 2020 Nov 11; 22(11): e23128. doi: 10.2196/23128. PMID: 33035175; PMCID: PMC7661105.

34.

Mohanty

Hughes

Salathé

. Using deep learning for image-based plant disease detection. Frontiers in Plant Science. 2016; 7: 1419.

35.

Menzies

Wald

Frank

. Data mining static code attributes to learn defect predictors. IEEE Transactions on Software Engineering. 2006; 33(1): 2-13.

36.

Stolfo

Fan

Lee

Prodromidis

Chan

. Cost-based modeling for fraud and intrusion detection: Results from the JAM project. In: Proceedings DARPA Information Survivability Conference and Exposition (DISCEX’00). IEEE; 2000 Jan 25. Vol. 2; pp. 130-144.

37.

World Health Organization. Ethics and governance of artificial intelligence for health. WHO guidance. 2021; Available from: https://www.who.int/publications/i/item/9789240029200.

38.

Perelman

Barrett

. The Mayfield handbook of technical and scientific writing. McGraw-Hill, Inc.; 1997 Jan 17.

39.

Mahato

Paul

. Analysis of region of interest (RoI) of brain for detection of depression using EEG signal. Multimedia Tools and Applications. 2023; na: 1-23. doi: 10.1007/s11042-023-15827-7.

40.

Metropolis

Rosenbluth

Teller

. Equation of state calculations by fast computing machines. The Journal of Chemical Physics. 1953 Jun 1; 21(6): 1087-92.

41.

Khadidos

Alyoubi

Mahato

Khadidos

Mohanty

. Computer Aided Detection of Major Depressive Disorder (MDD) using Electroencephalogram Signals. IEEE Access. 2023; 11: 41133-41141. doi: 10.1109/ACCESS.2023.3262930.

42.

Hanson

. Introduction to Bayesian image analysis. In: Proc SPIE 1898, Medical Imaging 1993: Image Processing. 1993. pp. 716-731. doi: 10.1117/12.154577.

A novel dataset and local interpretable model-agnostic explanations (LIME) for monkeypox prediction

Abstract

Keywords

1. Introduction

3. Methodology

3.1 Data collection

Table 2 Percentage of applied data augmentation

Table 4 Parameter used to identify super pixels

4.1 Experiment setup

4.2 Dataset

Table 5 Samples of images for dataset

Table 6 Results of performance metrics of studies one and two

Ethical approval

Funding

Data availability

Author contributions

Footnotes

Conflict of interest

References

Table 2
Percentage of applied data augmentation

Table 4
Parameter used to identify super pixels

Table 5
Samples of images for dataset

Table 6
Results of performance metrics of studies one and two