Automated Monkeypox Classification Using EfficientNetB3: A Deep Learning Approach for Multi-Class Skin Lesion Detection

The 2022 international outbreak of monkeypox highlighted critical deficiencies in rapid diagnostic capability, particularly in differentiating monkeypox from clinically similar viral exanthems. This study presents the first implementation of EfficientNetB3 with a two-stage transfer learning approach for four-class skin lesion classification (monkeypox, chickenpox, measles, and normal skin). Key methodological contributions include inverse-frequency class weighting to address extreme data imbalance (3.2:1 ratio), a combination of L2 regularization and progressive dropout (0.6→0.5→0.4→0.3), and a six-transformation data augmentation pipeline. Trained on only 770 images, the smallest dataset in comparative literature, the model achieved a validation accuracy of 91.56% (95% CI: 87.68%–95.44%), the highest reported performance for multi-class monkeypox classification. Per-class F1-scores demonstrate balanced minority-class learning: chickenpox (F1: 87.72%), measles (F1: 86.67%), monkeypox (F1: 91.59%), and normal skin (F1: 94.74%). A one-sample t-test against the ResNet50 5-fold cross-validation baseline (91.04% ± 1.71%) confirmed no statistically significant difference in overall accuracy (t = 0.30, p = 0.77), while EfficientNetB3 achieved notable improvements in minority-class performance (+4.95 percentage points (pp); chickenpox F1 +4.75 pp). EfficientNetB3 delivers these results with 48% fewer parameters, 40% less training time, and 13% faster inference, demonstrating strong feasibility for deployment in resource-limited clinical settings.