Purpose – To develop a lightweight and accurate computer-aided diagnostic framework for multiclass skill lesion classification with focus on Mpox that enables rapid and reliable detection during fast-spreading viral outbreaks, particularly in resource-constrained healthcare environments. Design/methodology/approach – The proposed framework integrates a Lightweight Depthwise-Separable Convolutional Neural Network with an Adaptive Bayesian Boosted Learning Module (LDSCNN–ABBLM). Contrast-limited adaptive histogram equalization (CLAHE) is applied to enhance lesion visibility, while class-weighted learning mitigates data imbalance. The LDSCNN backbone performs efficient feature extraction using depthwise-separable convolutions, and ABBLM employs Bayesian optimization via Optuna to adaptively tune boosting parameters over 15 trials. The model is evaluated on the MCSI and clinically validated MSLD v2 datasets and compared against multiple weighted baseline classifiers, including Weighted Random Forest, Weighted Linear SVM, Weighted LightGBM, Weighted Extra Trees, and Weighted XGBoost. Findings – The proposed model achieves validation accuracies of 0.9893 and 0.9983 on the MCSI and MSLD v2 datasets, respectively, demonstrating strong diagnostic reliability and superior generalization performance compared to all baseline models. Research limitations/implications – The model focuses exclusively on image-based diagnosis and does not incorporate clinical parameters such as patient history, symptoms, or laboratory findings, which are essential for comprehensive diagnosis. Furthermore, although the LDSCNN architecture is computationally efficient, real-time deployment on low-power edge devices in clinical settings may face challenges due to hardware variability and potential latency issues. Practical implications – The lightweight and scalable design enables deployment in low-resource and point-of-care settings, facilitating early Mpox detection and outbreak containment. Adherence to clinical validation and ethical standards further supports the framework’s integration into real-world AI-assisted dermatological and public health decision-support systems. Originality/value – This study introduces a unified lightweight diagnostic framework that synergistically combines depthwise separable convolutional feature extraction with adaptive Bayesian-optimized boosting, offering a novel balance between high diagnostic accuracy and computational efficiency for Mpox lesion classification.