An explainable deep learning and machine learning model for Mpox classification using dermatologic images

Monkeypox, now known as Mpox, has emerged as a significant global health concern. Since the onset of the 2023–2025 epidemic, over 118,000 confirmed cases have been reported across 127 countries, with more than 800 fatalities, primarily in the Democratic Republic of the Congo (DRC). In August 2024, the World Health Organization (WHO) declared the outbreak a public health emergency of international concern. The virus primarily spreads through close contact, with recent cases in Europe and America underscoring its growing global reach (World Health Organization, 2024; Weintraub, 2025).

Mpox, caused by the Mpox virus, presents a variety of symptoms in humans. It typically begins with flu-like signs, such as fever, headache, muscle aches, and fatigue. This is followed by a distinctive rash that starts on the face and spreads to other parts of the body, progressing from small, raised bumps to fluid-filled blisters (CDC, 2024; Usman et al., 2017). Other possible symptoms include swollen lymph nodes, sore throat, cough, and gastrointestinal issues like nausea, vomiting, and diarrhea (Thornhill et al., 2022). The incubation period ranges from 3 to 17 days, during which the individual remains asymptomatic (CDC, 2024).

Following the onset of Mpox symptoms, a timely and accurate diagnosis is crucial due to its similarity to other skin lesions. Traditional methods, including PCR, genome sequencing, antigen/antibody detection, and virus isolation, are accurate but resource-intensive, requiring specialized equipment and trained personnel, often causing delays, particularly in resource-limited or rural settings (Liu et al., 2024a). Machine learning offers a transformative solution by analyzing complex patterns in skin lesion images, aiding dermatologists in distinguishing Mpox from similar conditions. This enhances diagnostic accuracy, reduces human error, and supports personalized treatment. Given the limitations of conventional approaches, such as subjectivity, limited accessibility, and delayed results, researchers increasingly rely on image-based machine learning models for faster, more consistent, and scalable Mpox diagnosis.

Building on recent advances in image-based diagnostics (Debnath et al., 2025; Cao et al., 2025; Setegn & Dejene, 2025; Ali et al., 2024; Thieme et al., 2023), the accurate classification of Mpox requires the integration of cutting-edge machine learning and deep learning techniques within dermatologic image analysis. Beyond supporting dermatologists in diagnosis, these technologies offer the potential to improve patient outcomes by enabling timely, precise identification of Mpox, thereby reducing the risks of misdiagnosis and severe disease progression.

Hence, this study presents a novel framework integrating convolutional neural networks (CNNs) for feature extraction with a Bayesian-optimized Weighted C-SVM (BOW-SVM) classifier. The CNN captures discriminative patterns from dermatologic images, while the BOW-SVM improves classification performance by optimally tuning hyperparameters via Bayesian optimization and effectively handling class imbalance. Bayesian optimization (BO) was chosen for hyperparameter tuning due to its efficiency and systematic exploration of the search space. Unlike grid search, which exhaustively tests all combinations, or random search, which samples randomly. Studies (Liu et al., 2024b; Yang, Liu, & Wen, 2024; Yousaf et al., 2024; Lu, Zhang, Fan, Wan, & Luo, 2025; Li, Zhou, & Wang, 2025) have demonstrated the effectiveness of BO, validating its efficacy in various applications. This synergy improves accuracy, robustness, and generalization, making it particularly suited for complex clinical tasks. To the best of our knowledge, this integrated approach has not been employed in any previous study on Mpox prediction.

The main contributions of this study are as follows.

1. Contrast-limited adaptive histogram equalization (CLAHE) was employed to enhance the visibility and clarity of skin lesion features, thereby improving subsequent feature extraction and classification accuracy.

2. We designed a novel lightweight depthwise separable convolutional neural network (LDSCNN) to efficiently extract discriminative Mpox-specific features from enhanced images while minimizing computational complexity.

3. An adaptive Bayesian-optimized boosting learning machine (ABBLM) was proposed, where Bayesian optimization via Optuna was used to fine-tune key hyperparameters, improving classification performance and generalization.

4. This framework further enabled the analysis of parameter importance and performance history, offering valuable insights into the influence of hyperparameter tuning on model behavior and performance.

5. The proposed LDSCNN-ABBLM framework was rigorously compared with several state-of-the-art models in terms of classification accuracy, model complexity, and architectural efficiency, consistently demonstrating outstanding results.

This study advances Mpox diagnosis by introducing a robust and interpretable deep learning–machine learning framework tailored for real-world clinical deployment. The proposed system integrates CLAHE-based image enhancement, a lightweight depthwise separable convolutional neural network (LDSCNN) for efficient feature extraction, and an Adaptive Bayesian-Optimized Boosting Learning Machine (ABBLM) for accurate classification. This combination not only improves diagnostic performance but also addresses critical challenges such as overfitting and limited generalization. A total of 30 optimization trials were conducted using Bayesian optimization, enabling systematic hyperparameter tuning. The resulting insights into parameter importance and performance history contributed to enhanced model accuracy, stability, and decision reliability in classifying Mpox from dermatologic images.

The remainder of this study is organized as follows: Section 2 reviews related works, while Section 3 describes the system model framework and methodology. Section 4 reports the experimental results and comparative performance analysis. Finally, Section 5 concludes the study and outlines directions for future research.

Recent advances in machine learning and deep learning have shown promise in automating skin lesion analysis, enhancing diagnostic accuracy, and reducing the burden on healthcare professionals.

Convolutional Neural Networks (CNNs) are among the most powerful deep neural network architectures for processing image data (Albawi, Bayat, Al-Azawi, & Ucan, 2018). Convolutional, nonlinear, and pooling layers are some of the techniques CNNs use to extract key features from raw images automatically. Several computer vision tasks, such as image classification, object detection, image segmentation, and face recognition, have achieved notable successes (Li, Liu, Yang, Peng, & Zhou, 2021).

CNNs have also made notable progress recently in healthcare applications; for instance, Mujahid et al. (2024) propose EfficientNet for malaria detection using red blood cell images and achieve 97.57% accuracy. It includes performance comparisons with pre-trained models and uses k-fold cross-validation, demonstrating practical benefits for healthcare professionals. Majumdar, Pramanik, and Sarkar (2023) proposed an ensemble of three pre-trained CNN models, GoogleNet, VGG11, and MobileNetV3Small, to detect breast cancer in histopathological images. The models achieved classification accuracies of 99.16% and 96.95% on two benchmark datasets. Also, a study by Kumari and Ghosh (2023) presents a transfer learning-based AI system for classifying breast cancer from histopathological images using three deep CNN architectures: VGG-16, Xception, and DenseNet-201. It classifies images as benign or malignant regardless of the magnification factor. When evaluating the Invasive Ductal Carcinoma (IDC) and Breast Cancer Histopathological (BreaKHis) datasets, the system achieved a classification accuracy of 99.42% and 99.12%, respectively, outperforming state-of-the-art methods.

With the global emergence of Mpox, researchers have increasingly focused on leveraging deep learning to classify skin lesions associated with the disease. Gabriel, Kumar, and Prasad, (2025) proposed the D2T-PT model, integrating transfer learning and digital twin technology for Mpox detection. Using the MSLD dataset, their adaptive NasNetMobile model achieved 100% recall, 98% ROC, and 97.78% accuracy. Raha et al. (2024) employed an attention-based MobileNetV2 model for Mpox image classification, reporting outstanding performance compared to baseline models, with accuracies of 92.28% on the extended MSID dataset, 98.19% on the original MSID dataset, and 93.33% on the MSLD. Notably, their study avoided data augmentation techniques to maintain the authenticity of medical images, enhancing the model's ability to learn from real-world data. However, this decision may limit the model's robustness to variations and biases in clinical images. Al-Gaashani, Xu, and Obsie (2024) proposed an advanced Mpox detection model based on MobileNetV2, enhanced with progressive transfer learning, a Residual Dilated Spatial Pyramid Integration (ResDSPI) block, and an Efficient Channel Attention (ECA) module.

Karaddi, Sharma, and Bhattacharya (2024) developed the Softflatten network (Softflatten-Net), a deep convolutional neural network architecture designed for the automated detection and classification of monkeypox using skin lesion images. This model addresses challenges like gradient vanishing and overfitting and achieves high classification accuracy across multiple classification tasks. The proposed network outperforms existing models in terms of accuracy, sensitivity, and precision, and its performance is validated on several datasets. It also includes interpretability techniques to analyze misclassifications. Asif, Zhao, Li, Tang, and Zhu (2024a) introduced the Choquet Fuzzy Integral-based Ensemble (CFI-Net) for accurate classification of skin and oral diseases, with a focus on detecting Mpox, foot ulcers, and oral diseases. The approach enhances base classifiers through transfer learning and optimizes fuzzy measures using meta-heuristic algorithms, achieving remarkable classification accuracy across multiple disease detection tasks.

Maqsood, Damaševičius, Shahid, and Forkert (2024) proposed an integrated CAD system for monkeypox classification that incorporates fusion-based contrast enhancement, modified deep learning models, feature fusion using sparse decomposition, entropy-controlled firefly feature selection, and a multi-class SVM classifier. This approach demonstrated superior performance compared to state-of-the-art methods in both visual and quantitative evaluations, and Asif, Zhao, Li, Tang, and Zhu (2024b) introduced the CGO-Ensemble framework for Mpox detection by integrating transfer learning models with Chaos Game Optimization for weighted fusion.

Several studies have effectively employed pre-trained Convolutional Neural Networks (CNNs) for feature extraction in Mpox classification (Kesav & MG, 2023; Kesav & Jibukumar, 2022; Khairandish, Sharma, Jain, Chatterjee, & Jhanjhi, 2022; Zhang et al., 2021). Kesav and MG (Kesav & MG, 2023) used GoogLeNet for feature extraction from chest X-rays and classified them using various machine learning models. The Bayesian-optimized kernel SVM achieved top accuracy in both binary and three-class COVID-19 detection tasks. In a related study, (Kesav & Jibukumar, 2022) introduces a low-complexity framework for brain tumor analysis, employing RCNN with a two-channel CNN backbone for feature extraction and SVM for classification, achieving 98.21% accuracy and 98.83% confidence with reduced execution time across three tumor types. Similarly, a hybrid CNN-SVM framework that combines deep feature extraction with threshold-based segmentation for accurate detection and classification of brain tumors in MRI images was proposed by (Khairandish et al., 2022), achieving high performance and efficient tumor localization across multiple tumor types.

Effective preprocessing is essential in medical image analysis to enhance feature visibility, reduce noise, and improve lesion contrast before deep learning–based feature extraction. Contrast Limited Adaptive Histogram Equalization (CLAHE), introduced by Zuiderveld (1994), has been widely used to boost image clarity and deep learning performance across diverse diagnostic tasks. Yoshimi et al. (2024) demonstrated that CLAHE preprocessing significantly increases the robustness and accuracy of encoder-decoder CNNs for automated segmentation of temporomandibular joint articular disks in MRI datasets. Ji et al. (2024) further applied CLAHE within a multi-network fusion framework, enhancing dermoscopy-assisted diagnosis and differentiation of scalp psoriasis and seborrheic dermatitis. CLAHE has also been incorporated into deep learning pipelines for multi-class skin lesion classification, strengthening the discriminative power of global average pooling layers and improving lesion boundary visibility (Raghavendra, Charitha, Begum, & Prasath, 2023). Advanced CLAHE-based techniques, including Sailfish Optimizer enhanced U-Net and Bayesian optimization CLAHE, have shown substantial gains in segmentation and classification accuracy by optimizing local contrast and adaptive clip limits (Yogalakshmi & Rani, 2024; Han, Choi, Kim, & Lee, 2025).

As summarized in Table 1, existing Mpox detection frameworks face persistent challenges such as limited dataset diversity, inadequate attention to fine-grained lesion attributes, and insufficient optimization for real-time or mobile deployment. These limitations hinder model robustness and scalability, particularly under variations in lighting, skin tone, and environmental conditions. As a result, the performance of many existing methods may not generalize effectively to real-world clinical scenarios or diverse populations.

To address these shortcomings, this study proposes the LDSCNN-ABBLM, a lightweight two-stage framework designed for automated and efficient Mpox diagnosis. In the first stage, contrast-limited adaptive histogram equalization (CLAHE) enhances lesion visibility, while a Lightweight Depthwise Separable Convolutional Neural Network (LDSCNN) extracts discriminative features with reduced computational cost. The second stage employs an Adaptive Bayesian-Optimized Boosting Learning Machine (ABBLM), whose hyperparameters are optimized using Optuna across 30 trials for improved performance. Employing five-fold cross-validation ensures robustness and mitigates overfitting. The proposed LDSCNN-ABBLM achieves high accuracy, interpretability, and generalization, providing a scalable and clinically deployable solution for real-time Mpox detection across heterogeneous image domains.

The LDSCNN-ABBLM framework comprises three layers: (1) data sourcing for image aggregation, (2) a deep feature pipeline for preprocessing and extraction using LDSCNN, and (3) optimized model evaluation for training, tuning, and performance assessment, as shown in Figure 1.

The data source layer serves as the foundational stage of the LDSCNN-ABBLM framework, focusing on the acquisition and aggregation of diverse skin lesion images relevant to Mpox and related diseases. Let the initial dataset be represented as:

where x_i denotes the i-th input image and y_i its corresponding label, and N is the total number of raw samples collected from various public sources and other medical image repositories. These sources ensure the inclusion of varied demographics and lesion types, improving generalization.

To ensure data quality, a validation function V(x_i, y_i) is applied to filter out irrelevant or low-quality samples:

This filtering process removes duplicates, mislabelled, and low-resolution images, resulting in a curated dataset $Dclean$⁠, where V(x_i, y_i) was applied to both the Mpox Skin Lesion Dataset (MSLD) and the Mpox Close Skin Images (MCSI) dataset, whose authors had already performed essential cleaning and preprocessing to ensure data quality, with Contrast Limited Adaptive Histogram Equalization (CLAHE) further applied in this study to enhance local contrast and improve image quality.

In this layer, we perform essential preprocessing and feature extraction to prepare the curated dataset $Dclean$ for effective model learning. This layer includes three main steps: image enhancement, feature extraction using LDSCNN, and dataset splitting.

First, each image $xi∈Dclean$ undergoes contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE), represented as:

This enhances local contrast, making lesion patterns more distinguishable. The enhanced images are then resized to a fixed dimension (H, W, C) for uniform processing.

Next, feature extraction is performed using the Lightweight Deep Separable CNN (LDSCNN) as shown in Figure 2, a 19-layer compact convolutional neural network. The 19-layer depth was specifically chosen to balance model complexity and computational efficiency, providing sufficient capacity to capture discriminative patterns from dermatologic images while minimizing overfitting. Let f(⋅; θ) denote the LDSCNN model with parameters θ. The extracted feature vector z_i for image $xiclahe$ is computed as:

These feature vectors ${zi}i=1N′$ (where $N′=|Dclean|$⁠) capture spatial and semantic patterns necessary for classification.

Finally, the dataset is partitioned into training, validation, and testing sets:

This split ensures proper model evaluation and generalization during later stages.

The optimized model evaluation layer is the final stage of the LDSCNN-based classification framework, where the extracted features are used for training, tuning, and evaluating the predictive model. This layer employs a Bayesian Learning–assisted Kernel SVM classifier, combined with hyperparameter optimization and cross-validation, to ensure robust and reliable performance.

Let ${(zi,yi)}i=1N′$ denote the feature–label pairs from the training dataset $Dtrain$⁠, where z_i represents the LDSCNN-extracted feature vector and y_i the corresponding ground-truth label. The Kernel SVM learns a decision function of the form:

• where K(⋅, ⋅) is the selected kernel function, β_i are the learned support vector coefficients, and b is the bias term.

To prevent overfitting and ensure generalizability, 5-fold cross-validation is employed on $Dtrain$⁠, where the dataset is divided into 5 equal subsets. For each fold, 4 subsets are used for training and 1 for validation, rotating across all combinations.

Bayesian Optimization is used to tune hyperparameters such as learning rate, number of estimators, and tree depth. The objective is to minimize a loss function $L$ over the validation folds:

After selecting the optimal hyperparameters θ*, the final model is evaluated on the test set $Dtest$ using standard metrics such as precision, recall, F1-score, and confusion matrix analysis.

This section presents a comprehensive evaluation of the proposed LDSCNN–ABBLM model for Mpox lesion classification across two benchmark datasets. The experimental pipeline begins with CLAHE preprocessing to enhance lesion visibility, followed by the LDSCNN architecture for efficient and discriminative feature extraction. The extracted deep features are subsequently classified using ABBLM, where Bayesian optimization, implemented via Optuna over 15 trials, is employed to identify the optimal hyperparameter configuration that maximizes classification accuracy. For comparative analysis, the proposed framework is evaluated against several weighted ensemble and margin-based classifiers, including Weighted Random Forest (W-Random Forest), Weighted Linear SVM (W-L SVM), Weighted LightGBM (W-LightGBM), Weighted Extra Trees (W-Extra Trees), and Weighted XGBoost (W-XGBoost). The evaluation further encompasses diagnostic reliability analysis and ablation studies, collectively demonstrating the effectiveness, robustness, and generalization capability of the proposed model.

This study utilized two benchmark datasets: the Mpox Close Skin Images (MCSI), developed by Campana et al. (2024) dataset and the Mpox Skin Lesion Dataset (MSLD v2.0), developed by Ali et al. (2024). The MCSI dataset consists of 400 high-quality images collected from public sources and pre-processed through cropping and zooming to emphasize lesion regions. It is evenly divided into four classes—Mpox, Chickenpox, Acne, and Healthy—supporting the evaluation of machine learning models on smartphone-captured images. In contrast, the MSLD v2.0 dataset comprises 755 clinically validated images from 541 patients across six categories: Mpox, Chickenpox, Measles, Cowpox, Hand-Foot-Mouth Disease (HFMD), and Healthy. Featuring participants of diverse skin tones, sexes, and nationalities, MSLD v2.0 ensures fairness, robustness, and ethical compliance. Its adherence to strict clinical validation and ethical standards enhances reliability and real-world applicability in AI-driven dermatological research and diagnosis.

All images were resized to 224 × 224 pixels to meet the input requirements of the deep learning models. Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied to enhance local contrast and highlight lesion details. To further improve diagnostic precision and generalization, data augmentation was performed using the ImageDataGenerator by applying controlled rotations, horizontal flipping, spatial translations, zoom variations, and brightness adjustments to enhance image diversity and mitigate overfitting.

In our study, the classification of Mpox, T_p denotes the number of true positives or F_p the number of false positives, indicating correct diagnosis. In contrast, false positives F_p or false negative F_n indicates a misdiagnosis. Detailed equations of these metrics can be found in Table 2. To ensure reliable performance assessment, 95% confidence intervals (CIs) were computed for each metric across folds, providing a statistical measure of variability and stability, while these metrics, along with confusion matrix and ROC–AUC analyses, formed the basis for evaluating both cross-validation and final test performance.

Tables 3 and 4 report the 5-fold cross-validation results of the evaluated classifiers on the MCSI and MSLD v2 datasets, with performance measured in terms of accuracy, precision, recall, F1-score, and average inference time, all accompanied by 95% confidence intervals. On the MCSI dataset, the proposed model achieved the best overall performance, attaining an accuracy of 0.9893 (95% CI: 0.9872–0.9914), precision of 0.9894 (95% CI: 0.9873–0.9916), recall of 0.9893 (95% CI: 0.9872–0.9914), and F1-score of 0.9893 (95% CI: 0.9872–0.9914), thereby outperforming all baseline methods. Among the competing classifiers, W-L-SVM demonstrated strong predictive capability while achieving the lowest inference time, indicating superior computational efficiency, whereas ensemble-based approaches such as Random Forest and Extra Trees exhibited higher inference costs despite competitive accuracy. Similarly, on the MSLD v2 dataset, the proposed model consistently delivered the highest performance, recording an accuracy of 0.9828 (95% CI: 0.9781–0.9876), precision of 0.9831 (95% CI: 0.9786–0.9876), recall of 0.9828 (95% CI: 0.9781–0.9876), and F1-score of 0.9828 (95% CI: 0.9782–0.9875), substantially surpassing all baseline classifiers. Although W-L-SVM again offered faster inference, the proposed approach provided a superior balance between predictive performance and robustness. The consistently narrow confidence intervals across both datasets further validate the stability and reliability of the proposed model for Mpox lesion classification.

The confusion matrices in Figures 3 and 4 illustrate the classification performance of the proposed model across five cross-validation folds on the MCSI and MSLD v2 datasets, respectively. For the MCSI dataset (Figure 3(a)–(e)), the model exhibits excellent discriminative capability across the four classes—Acne, Chickenpox, Monkeypox, and Normal—with Acne and Normal showing near-perfect classification and minimal confusion between Chickenpox and Monkeypox. For the MSLD v2 dataset (Figure 4(a)–(e)), the model consistently achieves high accuracy, with predictions predominantly concentrated along the diagonal, reflecting strong and stable recognition across all six disease categories—Chickenpox, Cowpox, HFMD, Healthy, Measles, and Monkeypox—with only minor misclassifications observed between visually similar classes such as Monkeypox and Chickenpox. The consistent diagonal dominance across all folds in both datasets demonstrates the robustness and stability of the proposed model during cross-validation.

Following cross-validation, final evaluation on unseen test data assessed generalization, emphasizing Mpox class-specific precision, recall, and F1-score in line with the study objective, as presented in Table 5. cross both datasets, the proposed model consistently performed at par with or better than other benchmark classifiers. On the MSLD v2 dataset, the proposed model achieved a precision of 75%, recall of 91%, and an F1-score of 82%, closely matching or slightly exceeding other strong performers such as W-L-SVM. On the MCSI dataset, the proposed model reached a perfect recall of 100% with precision and F1-score of 95% and 98%, respectively, demonstrating superior balance and generalization compared to competing methods. These results underscore the robustness and discriminative effectiveness of the proposed approach in accurately identifying Mpox, even under unseen, real-world conditions.

The test set confusion matrices in Figure 3(f) and Figure 4(f) demonstrate the model's strong generalization across both datasets. On the MCSI dataset, predictions are highly accurate and consistent across all classes, with minimal misclassifications, highlighting the model's reliability in multiclass skin disease detection. Similarly, on the MSLD v2 dataset, the model maintains precise classification with only a few isolated errors, confirming its robust performance on unseen data and overall stability across diverse conditions.

The test set ROC curves in Figure 5 demonstrate strong class-wise discriminative performance of the proposed model. On the MCSI dataset in Figure 5(a), all classes achieve high AUC values, with curves approaching the top-left corner, indicating near-perfect separability. Similarly, on the MSLD v2 dataset as shown in Figure 5(b), the model maintains consistently high AUCs across all classes, reflecting reliable generalization on unseen data. Confusion matrices show minimal misclassifications mainly between visually similar classes, and ROC curves indicate near-perfect separability, confirming reliable differential diagnosis and strong clinical generalization.

The performance of the proposed model is attributed to its advanced architecture, which employs depthwise separable convolutions for efficient and deep feature extraction. This design allows the proposed method to capture complex patterns while maintaining computational efficiency, making it well-suited for datasets with intricate spatial features. Its ability to generalize effectively and balance precision and recall underscores its reliability and positions it as the most effective model for tasks requiring high accuracy and robustness, especially in demanding applications like medical imaging. In summary, the proposed model emerges as the top performer across most metrics, particularly excelling in accuracy, specificity, and AUC, making it the most reliable model among those evaluated.

The proposed LDSCNN–ABBLM framework demonstrates computational efficiency with average inference times of 1.3966 ms/image on MCSI and 5.3471 ms/image on MSLD v2, supporting real-time Mpox diagnosis in resource-constrained environments. Beyond computational aspects, the framework holds clinical significance by facilitating rapid and reliable lesion recognition, potentially aiding early diagnosis and reducing clinician workload. MSLD v2 adheres to rigorous clinical validation and ethical approval standards, ensuring relevance to real-world medical settings. Furthermore, the framework is designed for seamless integration with existing healthcare systems, while acknowledging deployment challenges such as ensuring data privacy, maintaining model robustness across diverse populations, and meeting regulatory compliance for medical AI systems.

To assess the contribution of each component in the proposed framework, an ablation study was performed on both the MCSI and MSLD v2 datasets using 5-fold stratified cross-validation. The complete model configuration integrates CLAHE preprocessing, depthwise-separable convolutional layers, ABBLM ensemble classification, Bayesian hyperparameter optimization, and class weighting. To evaluate the impact of individual components, the following variants were examined: (i) without (w/o) CLAHE preprocessing, (ii) without (w/o) depthwise-separable convolutions, only with standard convolutions, (iii) disabling (w/o) Bayesian optimization, and (iv) training without class weights. Each configuration was independently trained and validated under identical conditions, and the mean performance with 95% confidence intervals was reported for accuracy, precision, recall, and F1-score.

Tables 6 and 7 summarize the ablation study results on the MCSI and MSLD v2 datasets. Excluding any individual component consistently led to decreased performance, underscoring the contribution of each module. Omitting CLAHE preprocessing caused a slight drop in accuracy, highlighting its role in enhancing lesion contrast. Replacing depthwise-separable convolutions with standard convolutions marginally reduced performance, demonstrating the effectiveness of lightweight feature extraction. Disabling Bayesian optimization resulted in a modest decline, indicating its importance in selecting optimal hyperparameters. Removing class weights resulted in the most notable decrease, highlighting the benefit of addressing class imbalance. In all, the full LDSCNN–ABBLM model achieved the highest accuracy, precision, recall, and F1-score on both datasets, confirming that the integrated components collectively enhance generalization and diagnostic reliability.

Figure 6 illustrates the Optuna-based hyperparameter tuning outcomes for the MCSI and MSLD v2 datasets, encompassing both parameter importance analysis and optimization trajectories. For the MCSI dataset as shown in Figure 6(a), the kernel parameter emerges as the most influential hyperparameter, contributing the majority of the performance gain. This is followed by the regularization parameter C, while the degree and gamma parameters exhibit comparatively lower importance. This distribution indicates that kernel selection primarily governs the classifier's ability to model complex decision boundaries, whereas fine-grained adjustments to other parameters play a secondary role. The corresponding optimization history shown in Figure 6(b) demonstrates rapid convergence, with the best objective value attained within the initial trials and only marginal fluctuations thereafter, suggesting an efficient exploration of the search space and stable optimization behavior.

A similar pattern is observed for the MSLD v2 dataset as shown in Figure 6(c–d), where the kernel parameter again dominates the importance ranking with a substantially higher contribution than the remaining hyperparameters. The C, degree, and gamma parameters exhibit progressively diminishing influence on model performance. The optimization curve for MSLD v2 shows a sharp improvement in the early trials followed by prompt stabilization, confirming both fast convergence and robustness of the Optuna-driven tuning process. Overall, these results consistently emphasize the decisive role of kernel selection across both datasets, while also demonstrating the effectiveness and stability of the adopted hyperparameter optimization strategy.

Hyperparameter optimization was performed using Optuna's Tree-structured Parzen Estimator (TPE) sampler with 15 trials for each dataset. The search space encompassed key SVM parameters, including the regularization term C, kernel type, gamma, and polynomial degree, as detailed in Table 8. The optimal configurations summarized in Table 9 indicate that the RBF kernel consistently achieved superior performance across both datasets and its multiclass counterpart, underscoring its effectiveness in modeling nonlinear feature interactions. Furthermore, the selected C values of 27.3567 and 373.8548 reflect an appropriate trade-off between margin maximization and classification accuracy. The use of the auto setting for the gamma parameter enabled adaptive feature scaling, contributing to stable convergence and strong generalization performance across both datasets.

1. Comparison with existing monkeypox detection methods

In this study, both validation and test sets were utilized to ensure reliable model evaluation. The validation set facilitated hyperparameter tuning and model selection, whereas the independent test set was employed for the final performance assessment. To maintain a fair and consistent comparison with existing studies, the test results of the proposed model were reported, as they best represent the generalization capability on unseen data. Accordingly, Table 10 summarizes the comparative analysis of the proposed framework against recent state-of-the-art approaches based on key evaluation metrics.

Rapid computer-aided diagnosis of skin lesions is crucial during fast-spreading viral outbreaks. In this study, we present a Lightweight Depthwise-Separable CNN with an Adaptive Bayesian Boosted Learning Module (LDSCNN–ABBLM) for multiclass Mpox lesion detection. The model combines CLAHE-based image enhancement, class weighting, Bayesian optimization, and ABBLM-based ensemble learning. When evaluated on the MCSI and MSLD v2 datasets, the proposed model achieved validation accuracies of 0.9893 and 0.9983, with corresponding 95% confidence intervals of 0.9872–0.9914 and 0.9781–0.9876, respectively. On the test set, it attained a precision of 75% and a recall of 91% for MCSI, while achieving 95% precision and 100% recall on MSLD v2, indicating strong generalization performance and robust lesion discrimination capability. Notably, MSLD v2 adheres to rigorous clinical validation and ethical approval standards, ensuring real-world applicability in AI-driven dermatology. The proposed framework is also designed for seamless integration with healthcare systems while complying with relevant regulatory and ethical guidelines. By leveraging depthwise-separable convolutions and optimized preprocessing, the model achieves high performance while remaining lightweight, making it suitable for deployment in resource-limited environments.

Future research will emphasize computational optimization and clinical integration. Model quantization and pruning will be explored to further reduce latency and memory footprint. Expanding the dataset with diverse demographic and imaging conditions will improve robustness and fairness. Integration of explainable AI (XAI) mechanisms will enhance interpretability, while collaboration with clinical partners will support real-world validation and regulatory compliance. These efforts aim to advance the model toward scalable, trustworthy, and portable Mpox diagnostic systems for low-resource medical settings.