With the 2020 coronavirus pandemic having a profound impact across the globe, reports of the emergence of monkeypox in 2023 reveal the threat of another global virus (McCollum and Damon, 2014). Monkeypox (Mpox) is a disease caused by Mpox virus and is a viral zoonotic disease of orthopoxvirus, so it can be transmitted from animals to humans through direct close contact (Islam et al., 2022), as well as human-to-human transmission. Mpox was first discovered in 1958 in a monkey in a laboratory in Copenhagen, Denmark (Ladnyj et al., 1972) and is known as Mpox due to its similar outbreak symptoms to smallpox.

The Mpox virus caused the first infection in the Congo in 1970. Since then, most cases have occurred in Congo, Central and West Africa, and the number of cases has gradually increased, affecting many people living near tropical regions. As of 2022, the World Health Organization(WHO) reports that several other non-African countries such as Europe and the United States have also reported cases of Mpox virus infection (Alakunle et al., 2020).

Since the declaration of the eradication of smallpox in 1980 and the subsequent cessation of smallpox vaccination, monkeypox has emerged as the predominant orthopoxvirus. Its symptoms resemble those of smallpox, thus garnering attention in the field of public health (Mohbey et al., 2022). In 2003, the United States became the first country outside Africa to experience a monkeypox outbreak. According to reports, in September 2018, a Nigerian tourist was infected with monkeypox in Israel; in September 2018, December 2019, May 2021, and May 2022, cases of infection were also reported in Singapore; while in May 2019, July 2021, and November 2021, cases of monkeypox were recorded in the United States. These countries may be located in Southeast Asia. In May 2022, a significant number of monkeypox cases were reported in some countries where the disease does not typically occur (Mohbey et al., 2022). According to the U.S. Centers for Disease Control and Prevention (CDC), as of December 21, 2022, Mpox cases have been reported in 94 countries worldwide, totaling approximately 83,424 cases. Due to the dire impact of the COVID-19 pandemic, Mpox cases have begun to be closely monitored, showing signs of potential epidemic transmission even though large-scale transmission has not yet occurred (Rogers et al., 2008; Alakunle et al., 2020; Gürbüz and Aydin, 2022; Sharif et al., 2022). As a result, deep anxiety and worry among people are gradually spreading.

Monkeypox virus infection is generally divided into two stages: the invasive stage and the skin rash stage. Symptoms during the invasive stage include fever, severe headache, swollen lymph nodes, back pain, muscle aches, and weakness. During the skin rash stage, a rash appears 1–3 days after the onset of fever, concentrating on the face and limbs (Adalja and Inglesby, 2022). The rash progresses from macules to papules, vesicles, pustules, forms scabs, and eventually falls off. These skin lesions are typically quite painful. When the rash appears, the patient becomes contagious. Monkeypox virus can spread through contact with infected individuals or animals. Specifically, when individuals come into contact with the ulcers, scabs, respiratory droplets, or oral fluids of an infected person, it may lead to the transmission of the disease to others (Simpson et al., 2020). Therefore, timely diagnosis of this disease is essential. According to the guidelines proposed by the WHO, healthcare personnel should wear protective gear when caring for patients. Additionally, patients need to be isolated, and they should maintain distance from others. Mpox exhibits subtle differences from other viruses such as smallpox, chickenpox, and measles, primarily in the inflammatory and rash symptoms induced within the human body. Apart from polymerase chain reaction (PCR), which is an effective diagnostic method (Ibrahim et al., 2021), non-specialists may find it challenging to differentiate them visually. Moreover, PCR testing is costly and typically requires a considerable amount of time to yield results. Therefore, it has not been widely adopted, posing challenges to rapid diagnosis (Reed et al., 2004). In order to involve non-specialists in the prevention and control of Mpox virus, researchers are actively seeking effective methods utilizing artificial intelligence to identify cases of Mpox. They are engaged in data collection and research experiments to gain a deeper understanding of this disease (Banerjee et al., 2022; Khafaga et al., 2022a). The development of automated identification algorithms could not only aid in the rapid identification of Mpox virus cases but also be utilized for training healthcare professionals who are not specialized in Mpox.

Despite the relatively low fatality rate of Mpox within the range of 1–10% (Gong et al., 2022), the absence of an antiviral therapy for curing monkeypox (Reynolds et al., 2017) underscores the importance of early detection in preventing its spread. Early identification plays a crucial role in the prevention, diagnosis, and treatment of this disease. With the rapid advancement of artificial intelligence models in the field of medicine, deep learning models for medical image analysis have been proposed for various medical science applications. However, there still exist numerous traditional manual classification methods for identifying Mpox, which suffer from apparent inefficiencies and high costs. The reliance on manual diagnostic classification is susceptible to human factors, leading to diagnostic variability and delayed confirmation of Mpox, depriving patients of timely access to appropriate treatment plans. Therefore, there is an urgent need to introduce advanced technologies to replace simplistic manual classification methods, thereby enhancing the chances of patient recovery and reducing the risk of transmission.

Over the years, deep learning (DL) has achieved remarkable success, exerting a profound impact on the core concepts of machine learning (ML) and artificial intelligence (AI). DL methods have demonstrated outstanding results in various industrial domains, overcoming limitations of traditional approaches. They have become powerful tools in the fields of image analysis and pattern recognition, showing extensive potential applications in disease detection. In particular, Convolutional Neural Networks (CNNs) have emerged as a cornerstone in image recognition, owing to their exceptional capabilities in feature extraction and non-linear representation. CNN, a DL neural network architecture, is commonly employed with image data as input. Through a series of operations, it extracts crucial features and information from input images, facilitating tasks such as classification or other related objectives. Some scholars have made significant progress by applying deep learning techniques to disease classification tasks. Sandeep et al. proposed a low-complexity CNN for identifying skin diseases such as psoriasis, melanoma, lupus, and chickenpox (Sandeep et al., 2022). Their research found that using existing VGGNet and image analysis techniques could accurately detect 71% of skin diseases. However, their proposed solution demonstrated optimal performance with an accuracy of approximately 78%. Glock et al (Glock et al., 2021). utilized the ResNet-50 model to develop a transfer learning approach for measles detection, exhibiting good performance on multiple rash image datasets with a sensitivity of 81.7%, specificity of 97.1%, and accuracy of 95.2%. Velasco et al (Velasco et al., 2019) introduced an intelligent smartphone skin disease recognition method based on MobileNet, reporting an accurate detection of chickenpox symptoms with an accuracy of about 94.4%. Sahin et al. (Sahin et al., 2022) developed a mobile application on the Android platform that rapidly diagnoses Mpox patients using deep learning technology, achieving an image classification accuracy of 91.11%.

Some scholars have already applied deep learning to the task of Mpox classification, achieving significant results (Mehrotra et al., 2020; Ahsan et al., 2023; Almufareh et al., 2023). A portion of researchers has also conducted relevant studies in the field of Mpox disease identification, as elucidated in the “Literature Review.” The main objective of this study is to determine and validate the optimal performing model for Mpox classification through transfer learning methods and classification models, with the aim of determining the incidence rate of Mpox. This study introduces a Mpox classification algorithm named MpoxNet, based on ConvNext (Liu et al., 2022), designed for the task of Mpox disease classification. Improvements to the ConvNext model are as follows: firstly, the introduction of a new dual-branch deep residual Squeeze and Excitation (D²RSE) module replaces the original ConvNext. This module has a dual-channel structure, significantly enhancing the model’s classification accuracy while achieving better performance with a reduced number of model parameters. By integrating the Convolutional Block Attention Module (CBAM) with ConvNext, we capture the spatial relationships of Mpox features during training, adaptively weighting features based on their importance in different spatial and channel dimensions, and suppressing irrelevant regions, allowing the model to selectively focus on features of different disease categories. For example, Sitaula et al. (Sitaula and Hossain, 2021)proposed a deep learning model based on attention, which employed the attention mechanism of VGG-16, aiming to more accurately capture the spatial relationships of key regions in chest X-ray images. In this experiment, MpoxNet is compared with other mainstream algorithm models, demonstrating higher classification accuracy in Mpox classification scenarios. The main contributions of this paper are as follows:

This paper is organized into the following sections: The second section outlines existing methodologies in Mpox image classification. The third section provides insights into the Mpox dataset and introduces the proposed enhancements. The fourth section details the performance evaluation and experimental results. Finally, the fifth section offers a comprehensive summary of the paper.

Skin lesions are prevalent in clinical practice, making precise detection and diagnosis crucial for accurate patient treatment. In recent years, the emergence of machine learning technologies has provided significant potential to assist in the identification and clinical decision-making for skin lesions (Fraiwan and Faouri, 2022). In this section, we will present various artificial intelligence techniques, including machine learning, CNNs, and transfer learning algorithms, applied for the detection and diagnosis of Mpox virus.

Iftikhar et al. (Iftikhar et al., 2023) proposed a novel filtering and ensemble technique for the rapid and accurate prediction of Mpox cases. This approach generates two subsequences, long-term trend sequences, and residual sequences, through filtering, and employs five machine learning models for prediction. Mandal et al. (Mandal et al., 2022) introduced a clustering method for Mpox cases that combines machine learning and Particle Swarm Optimization (PSO). Bhosale et al. (Bhosale et al., 2022) conducted similar work, utilizing linear regression, decision trees, random forests, elasticNet, and ARIMA for the prediction of Mpox cases. They further introduced a dataset named Mpox Skin Image Dataset (MSID), widely employed in various studies. For instance, Khafaga et al. (Khafaga et al., 2022b) introduced a novel framework for the classification of Mpox disease images. They employed the Random Fractal Search (BERSFS) using the Al-Biruni Earth Radius (BER) optimization method for fine-tuning on deep CNN layers. Saleh and Rabie et al. (Saleh and Rabie, 2023) proposed a Human Mpox Diagnosis (HMD) strategy based on artificial intelligence technology. This strategy comprises two key components: utilizing Improved Binary Chimpanzee Optimization (IBCO) and selecting features with significant value for transfer to the diagnostic model of ensemble learning. Ultimately, HMD achieved an accuracy of 0.98. Ahsan et al. (Ahsan et al., 2023)developed a Mpox diagnostic model using the Generalization and Regularization-based Transfer Learning Approach (GRA-TLA) for binary and multiclass classification. They tested ten different Convolutional Neural Network (CNN) models, including both binary and multiclass tasks, in three independent studies. In studies one and two, their model combined with Extreme Inception (Xception) achieved accuracies ranging from 77% to 88% in distinguishing individuals with and without Mpox. In study three, the accuracy using the ResNet-101 network ranged from 84% to 99%. Kumar et al. (Kumar, 2022) employed skin images for Mpox disease diagnosis, using various CNN models and machine learning algorithms. They extracted image features using Vgg16Net and AlexNet and applied classifiers such as Naive Bayes, Decision Trees (DT), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest. The Naive Bayes algorithm combined with Vgg16Net achieved the highest accuracy at 91.11%.

Research on skin lesion image classification through the combination of CNN and transfer learning methods has yielded significant benefits. Ahsan et al. (Ahsan et al., 2022) conducted an initial investigation into Mpox diagnosis. The authors collected images of patients infected with Mpox from various accessible portals and proposed a VGG16-based model for Mpox diagnosis. They constructed their dataset by gathering images from Google and utilized transfer learning to create a model based on the VGG16 architecture in two separate studies. The first study successfully distinguished Mpox from chickenpox, achieving an accuracy of up to 0.97, while the second study differentiated Mpox from other diseases (chickenpox, measles, and normal skin) with an accuracy of 0.89. Meena et al. (Meena et al., 2023) proposed a hybrid technique based on Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs) to embed knowledge graphs into various healthcare applications, providing enhanced data representation and knowledge inference. Experimental results demonstrate that the proposed model achieves an accuracy of 94% on the Mpox dataset. Bala et al. (Bala et al., 2023) presented a Convolutional Neural Network (MonkeyNet) based on an improved DenseNet-201 for Mpox image recognition. They evaluated its performance using the original images from the MSID dataset for training. Their model accurately identified Mpox on both the original and augmented datasets, achieving accuracies of 93.19% and 98.91%, respectively. Jaradat et al. (Jaradat et al., 2023) evaluated five pre-trained models, including VGG19, VGG16, ResNet50, MobileNetV2, and EfficientB3. Experimental results demonstrated that MobileNetV2 performed the best, achieving an accuracy of 0.98. Model validation across different datasets confirmed MobileNetV2’s highest accuracy of 0.94. Altun et al. (Altun et al., 2023) employed a similar approach and developed a hybrid function learning model incorporating hyperparameter optimization. They utilized custom models, including MobileNetV3-s, EfficientNetV2, ResNet50, Vgg19, DenseNet121, and Xception. Notably, the optimized MobileNetV3-s model exhibited the best performance, achieving an accuracy of 0.96. Uzun Ozsahin et al. (Uzun Ozsahin et al., 2023) applied deep learning models such as AlexNet, VGG16, and VGG19 for the detection task on Mpox and chickenpox datasets. Through their research methodology, they achieved a highest precision of 0.99. Sitaula et al. (Sitaula and Shahi, 2022) utilized deep learning techniques for Mpox diagnosis. They compared 13 pre-trained deep learning models, ultimately selecting the most outstanding model to build their system. The results indicated an accuracy of 0.87 for their Mpox diagnostic model. Ali et al. (Ali et al., 2022) created a dataset named “Monkeypox Skin Lesion Dataset (MSLD)” designed for automatic detection of Mpox disease from skin lesions. The images were primarily sourced from websites, news portals, and publicly available case reports. They employed pre-trained models such as VGG-16, ResNet50, and InceptionV3 for Mpox classification. To enhance the accuracy of Mpox detection, Sahin et al. employed six deep learning models, including ResNet-18, MobileNet, NasNetMobile, GoogLeNet, EfficientB0, and ShuffleNet, to distinguish Mpox images from images depicting other diseases. Among these models, MobileNet exhibited the best performance, achieving a 91.11% accuracy in image classification. Additionally, Almufareh et al. (Almufareh et al., 2023) utilized various independent CNN architectures to differentiate Mpox from non-Mpox cases. They validated their models using MSID and MSLD datasets. Javelle et al. (Javelle et al., 2023) redefined the emerging Mpox disease and designed a self-management questionnaire for case management, contact monitoring, and support for clinical research. Haque et al. (Haque et al., 2022) employed five deep learning models, VGG19, Xception, DenseNet121, EfficientNetB3, and MobileNetV2, and integrated spatial attention mechanisms for accurate classification of human Mpox. Yasmin et al. (Yasmin et al., 2022) obtained Mpox images from the Kaggle global dataset and used nine models for predictions. Among them, the optimal predictive model achieved an MSE value of 41922.55, R2 of 0.49, MAPE of 16.82, MAE of 146.29, and RMSE of 204.75. Meena et al. (Meena et al., 2024) established a deep learning model based on transfer learning to assist in diagnosing whether patients have Mpox. In the experiment, the InceptionV3 model they used achieved an accuracy of 98%.

Table 1 summarizes the relevant studies on Mpox diagnosis. However, research on Mpox diagnosis using deep learning remains relatively limited. Some investigations have assessed the potential of deep learning algorithms in identifying this disease. While the results suggest that deep learning could be a valuable tool for the diagnosis and control of Mpox, further research is needed to validate these findings and establish practical applications in real clinical settings.

This section encompasses four key stages: data collection, image preprocessing, model selection, and model optimization. In the first stage, we utilized the Monkeypox Skin Image Dataset (MSID), which is freely accessible on the Kaggle platform (Monkeypox Skin Images Dataset (MSID), n.d).. And due to limited data, we employed data augmentation techniques to generate additional images. During the data preprocessing phase, the collected images underwent operations such as resizing, normalization, and data augmentation, which are crucial for enhancing model performance. We selected eight commonly used models (SqueezeNet, ResNet18, ResNet34, ResNet50, Vgg16, DenseNet121, Swin-Tiny, and ConvNext-Tiny) for comparison to improve the accuracy of the Mpox virus detection model. In the model training phase, the selected models were trained using the preprocessed images, optimizing model performance by providing images and adjusting parameters. Finally, in the evaluation stage, metrics such as accuracy, precision, recall, and F1 score were employed to assess the models, with the best-performing model selected as the final model. Therefore, this approach utilizes deep learning techniques to analyze Mpox images, accurately classifying and diagnosing the disease based on visual features. The algorithm demonstrates high accuracy in disease classification, providing a valuable tool for the rapid and precise diagnosis of Mpox in clinical settings. The proposed method comprises multiple stages and processes, illustrated in Figure 1 , which encompass various steps and operations.

In Table 1 , despite slight variations in experimental parameters and training datasets compared to other methods, MpoxNet attains the highest recognition accuracy. Meanwhile, Table 2 illustrates that MpoxNet outperforms other networks with the lowest Flops and model parameter count, registering at 23.44G and 0.73M respectively. This allows the MpoxNet to not only serve as a real-time assessment tool but also be easily applied on smartphones for real-time identification and prediction of monkeypox cases. While MpoxNet demonstrates excellent performance in identifying monkeypox lesions, its precision in localizing widely scattered cases of monkeypox is somewhat lacking. In Figure 10 , although the heatmap can encompass the entire infected area, it fails to accurately mark the positions of individual infection points, highlighting the need for further improvement in this aspect.

This paper proposes a high-precision lightweight classification network, MpoxNet, based on ConvNext, for the diagnostic classification of Mpox. Firstly, a dual-branch depth separable convolution residual Squeeze and Excitation (D2RSE) module is designed. Then, CBAM is introduced to improve diagnostic accuracy and significantly reduce the model’s parameter count. Our proposed model is compared with SqueezeNet, ResNet18, ResNet34, ResNet50, VGG16, DenseNet121, Swin-Tiny, and ConvNext-Tiny. The proposed model achieves the highest accuracy, precision, recall, and F1-score among all tested models. In ablation experiments, the effectiveness of the D2RSE and CBAM modules is individually verified. By comparing MpoxNet with baseline networks from related papers using the BUSI dataset, the results clearly demonstrate the significant advantages of MpoxNet in terms of performance. This series of comparative experiments validates the crucial roles of the D2RSE and CBAM modules, providing solid evidence for the performance superiority of MpoxNet.

If the results of this study can be implemented, healthcare professionals may be able to improve patient prognosis and reduce medical costs. The research on deep learning and transfer learning techniques for automated Mpox detection not only provides innovative approaches for disease diagnosis but also opens new avenues for ad-dressing diagnostic challenges of other infectious diseases. Breakthroughs in this field will profoundly impact the medical domain, paving the way for the development of future diagnostic tools and methods.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

The manuscript presents research on animals that do not require ethical approval for their study. Written informed consent was obtained from the minor(s)’ legal guardian/next of kin for the publication of any potentially identifiable images or data included in this article.

JS: Software, Visualization, Writing – original draft, Writing – review & editing. BY: Methodology, Supervision, Writing – review & editing. ZS: Validation, Writing – review & editing. JZ: Visualization, Writing – review & editing. YD: Resources, Writing – review & editing. YG: Data curation, Writing – review & editing. YC: Data curation, Writing – review & editing.