Improved EfficientNet-B4 for Melanoma Detection

Abstract

Skin cancer is a deadly disease. Early detection of skin cancer can reduce the cost of treatment. Despite the low incidence, melanoma can cause 75% of skin cancer deaths. With the rapid development of computer vision, it is necessary to develop an image-based melanoma detection tool to assist doctors in pathological detection. In this paper, we propose an automatic detection method for melanoma based on EfficientNet-B4. This method is integrated into the medical visual assistance system to quickly and effectively provide auxiliary treatment to patients. Moreover, We used network fine-tuning technology to adapt the model to the dermoscopic lesion image data. We evaluated the experimental results on the ISIC 2020 challenge dataset collected by the International Skin Imaging Collaboration, which contains the largest publicly available collection of dermoscopic images of skin lesions. The experimental results indicate that our method has better performance in AUC-ROC score compared to vgg-16 and ResNet-50.