The Effects of L2 Regularization in EfficientNet for Human Skin Disease Multi-Class Classification

Skin diseases are complex problems that can have serious consequences for human health, including the risk of skin cancers such as melanoma. Identification and classification of skin diseases require extensive experience and knowledge in the field of dermatology. Therefore, this study aims to develop an early detection system using deep learning techniques with a pre-trained EfficientNet model to classify human skin disease images. In this study, we used two datasets, Dermnet and ISIC 2019, which contain images of different types of human skin diseases. We performed augmentation and oversampling on the datasets to overcome data imbalance and complexity. Then, we trained the EfficientNet model without and with Ridge Regression (L2) regularization on both datasets. The experimental findings suggest that using L2 regularization enhances model performance by lowering the rate of overfitting, resulting in a higher f1-score. The EfficientNet model with L2 regularization achieved an f1-score of 81.3% for the ISIC 2019 dataset, and an f1-score of 61.4% for the Dermnet dataset.