From survey to solution: A deep learning framework for reliable monkeypox diagnosis using skin images

Monkeypox, a re-emerging zoonotic disease, poses a global health threat due to its rapid transmission and visual similarity to other skin lesions such as chickenpox, measles and acne. Deep learning methods, which detect monkeypox from skin images, offer a promising solution to overcome the limitations of manual and PCR-based diagnoses, which are time-consuming, error-prone and impractical in low-resource settings. However, existing methods are limited by poor dataset quality, weak generalizability and inconsistent benchmarking. Practical issues, such as lesion variability, image noise, unsuitable augmentations and minimal preprocessing, pose further challenges to clinical deployment. This study addressed these issues through a three-fold contribution: a comprehensive survey of deep learning methods analyzing their strengths and limitations; the development of a diverse, clinically representative benchmark dataset to better assess model generalizability; and a robust deep learning ensemble framework that improved diagnostic accuracy across diverse skin images. The proposed ensemble method incorporated practical classes, a noise-free and balanced dataset, clinically relevant augmentations and effective preprocessing steps, achieving over 95% accuracy across major public datasets to ensure robustness and readiness for clinical deployment. Explainability analysis using Shapley Additive exPlanation (SHAP) confirmed the method’s reliability across all skin tones and body parts. A paired t-test showed that the ensemble model performed significantly better than individual models across four public datasets (p=0.005, Cohen’s d=2.38).