The L2-EffCANet: A Novel Overfitting-Resistant EfficientNetV2S with Attention Mechanism and L2 Regularization for Skin Disease Classification

Abstract

Skin diseases are a common health problem that is often underestimated. However, some types of skin diseases can become cancerous and fatal if not treated properly, such as melanoma. Melanoma is caused by excessive exposure to ultraviolet rays and has a 99% cure rate if diagnosed early, but this figure drops to 20% in advanced stages. In developing countries, uneven distribution of medical personnel and geographical challenges lead to many skin diseases going undiagnosed. This study develops a multi-class model for classifying skin diseases using transfer learning, leveraging pre-trained models like EfficientNetV2S to address overfitting and improve accuracy. The model is trained on a dataset from DermNet consisting of 23 classes and a total of 19,559 skin disease images. Data augmentation is used to reduce class imbalance. The EfficientNetV2S model with the addition of CA and L2 regularization at the end of the model architecture achieves a test accuracy of 71.78% and demonstrates stable superiority, surpassing previous research. The study shows that deep learning can help with the early detection of skin diseases, thus improving healthcare services.