Improving Skin Disease Classification with EfficientNetB0: A Deep Learning Approach to Overfitting Control

This study presents a deep learning approach for skin disease classification using the EfficientNetB0 architecture, targeting five distinct categories: Normal, Melanoma, Arsenic, Psoriasis, and Eczema. Using the EfficientNetB0 model, The model was trained on a large dataset of 7,356 dermatoscopic images, all meticulously annotated with professional diagnoses. To improve the generalization capabilities of the model, a number of data augmentation techniques were employed. To improve generalization, data augmentation was applied and conducted experiments by training the model over different duration with varying epochs (20, 30, and 50) to identify the most effective training period. The best performance was achieved at 50 epochs, with a training accuracy of 0.9988 and a validation accuracy of 0.8912, demonstrating strong capability in identifying skin diseases. This performance illustrates the robust ability of the model to accurately identify and classify skin conditions. However, some classification errors were observed, particularly between conditions with similar visual characteristics, such as Psoriasis and Eczema. Overfitting was mitigated using data augmentation, L2 regularization, and dropout, though further enhancements such as adaptive learning rates or early stopping could improve performance. This EfficientNetB0-based model shows substantial potential as a supportive tool in dermatological diagnostics, providing a solid foundation for automated skin disease classification with opportunities for future optimization.