Influence of activation function in deep learning for cutaneous melanoma identification

Abstract

Malignant melanoma is an aggressive skin cancer requiring early detection for effective treatment. In this study, it is hypothesized that the choice of activation function affects the classification performance of pre-trained models in melanoma detection, and that the optimal activation function varies across deep CNN architectures. The impact of various activation functions (ReLU, LeakyReLU, ELU, GELU, Swish, Mish, PReLU) on the diagnostic accuracy of ResNet152, DenseNet201, and EfficientNet-B4 models was investigated. The study was conducted using a combined ISIC dataset, comprising dermoscopic images collected between 2018 and 2020. Findings indicate EfficientNet-B4 with LeakyReLU achieved the highest accuracy of 90.5%, while DenseNet201 benefited most from ReLU (90.3%). Results confirm the influence of activation function selection, demonstrating architecture-specific optimal choices for enhanced classification.

Ähnliche Arbeiten