MobileNetV2 and EfficientNetB3 Ensemble for Scalable Dermoscopic Skin Lesion Classification.

Background: Skin diseases, ranging from mild benign lesions to life-threatening malignant conditions, remain a major global health concern. Early and accurate diagnosis is critical to avoid complications, yet this is often limited by the shortage of dermatologists and the subjective nature of visual inspections, particularly in low-resource settings. To address this challenge, this study proposes an automated deep learning framework for skin disease classification using dermoscopic images. Methods: The framework employs a hybrid learning approach by integrating transfer learning and ensemble learning techniques. Specifically, MobileNetV2 and EfficientNetB3 models were combined to leverage their unique strengths, thereby enhancing generalization and predictive accuracy. The system was trained on a well-annotated dataset of 22,177 dermoscopic images, representing eight diagnostic categories that include benign, malignant, and pre-cancerous skin conditions. Results: Experimental results demonstrated strong classification performance, achieving a training accuracy of 96.81%, validation accuracy of 87.66% (loss of 0.455), and test accuracy of 86%. To improve clinical trust and interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to highlight the image regions that contributed most to the model’s decisions. In addition, a user-friendly diagnostic interface was developed, enabling real-time image input, automated analysis, and clear interpretive guidance. This makes the system accessible not only to healthcare providers but also to non-specialists, bridging gaps in dermatological care. Conclusion: The proposed solution offers a reliable, interpretable, and scalable application of artificial intelligence for skin disease screening, with significant implications for tele-dermatology and seamless integration into clinical workflows. DOI: https://doi.org/10.59564/amrj/03.04/014