DeepGenMon: A Novel Framework for Monkeypox Classification Integrating Lightweight Attention-Based Deep Learning and a Genetic Algorithm

Background: The rapid global spread of the monkeypox virus has led to serious issues for public health professionals. According to related studies, monkeypox and other types of skin conditions can spread through direct contact with infected animals, humans, or contaminated items. This disease can cause fever, headaches, muscle aches, and enlarged lymph nodes, followed by a rash that develops into lesions. To facilitate the early detection of monkeypox, researchers have proposed several AI-based techniques for accurately classifying and identifying the condition. However, there is still room for improvement to accurately detect and classify monkeypox cases. Furthermore, the currently proposed pre-trained deep learning models can consume extensive resources to achieve accurate detection and classification of monkeypox. Hence, these models often need significant computational power and memory. Methods: This paper proposes a novel lightweight framework called DeepGenMonto accurately classify various types of skin diseases, such as chickenpox, melasma, monkeypox, and others. This suggested framework leverages an attention-based convolutional neural network (CNN) and a genetic algorithm (GA) to enhance detection accuracy while optimizing the hyperparameters of the proposed model. It first applies the attention mechanism to highlight and assign weights to specific regions of an image that are relevant to the model’s decision-making process. Next, the CNN is employed to process the visual input and extract hierarchical features for classifying the input data into multiple classes. Finally, the CNN’s hyperparameters are adjusted using a genetic algorithm to enhance the model’s robustness and classification accuracy. Compared to the state-of-the-art (SOTA) models, DeepGenMon features a lightweight design that requires significantly lower computational resources and is easier to train with few parameters. Its effective integration of a CNN and an attention mechanism with a GA further enhances its performance, making it particularly well suited for low-resource environments. DeepGenMon is evaluated on two public datasets. The first dataset comprises 847 images of diverse skin diseases, while the second dataset contains 659 images classified into several categories. Results: The proposed model demonstrates superior performance compared to SOTA models across key evaluation metrics. On dataset 1, it achieves a precision of 0.985, recall of 0.984, F-score of 0.985, and accuracy of 0.985. Similarly, on dataset 2, the model attains a precision of 0.981, recall of 0.982, F-score of 0.982, and accuracy of 0.982. Moreover, the findings demonstrate the model’s ability to achieve an inference time of 2.9764 s on dataset 1 and 2.1753 s on dataset 2. Conclusions: These results also show DeepGenMon’s effectiveness in accurately classifying different skin conditions, highlighting its potential as a reliable and low-resource tool in clinical settings.