Deep Learning-Based Monkeypox Detection: A Hybrid Approach Using DenseNet121 and MobileNetV2

Abstract

Due to recent outbreaks outside of endemic areas, Monkeypox is a newly emerging zoonotic disease that has drawn international attention. This study utilized two publicly available collections—the Monkeypox Skin Lesion Dataset (MSLD) and its updated version, MSLDV2.0, which consist of 2607 and 10572 clinical images, respectively. Clinical photographic images of confirmed Monkeypox lesions and comparative non-Monkeypox dermatoses. Early and precise lesion detection is therefore crucial for effective containment and treatment. We propose a hybrid deep-learning approach that fuses the hierarchical feature extraction capabilities of DenseNet121 with the computational efficiency of MobileNetV2 for reliable Monkeypox identification. On MSLDV2.0, our model achieved 98 % accuracy, and on MSLD, it reached 99.18 %. The confusion matrices confirm robust discrimination between Monkeypox and non-Monkeypox classes, outperforming existing methods. A key limitation of this work is the moderate size and demographic homogeneity of the datasets, which may not fully capture real-world variations in skin tone or lesion presentation. Future research should incorporate larger, multi-center image repositories, evaluate performance across diverse populations, and assess real-time deployment in resource-constrained clinical settings.