Identifying Melanoma Using Transfer Learning and Convolutional Neural Networks : An investigation of skin disease pre-training

Abstract: Melanoma, the deadliest form of skin cancer, has become an increasingly common and pressing health issue. Early detection and treatment can be life-saving, but also poses a challenge. A possible solution presents itself in Convolutional Neural Networks (CNNs). When utilizing transfer learning these algorithms show great potential as a means for identifying melanoma. This study investigates the use of a skin disease pre-trained CNN in a binary skin cancer classification task. This was done by training and comparing the performance of four EfficientNet-B0 CNN models on the SIIM-ISIC 2020 Challenge Dataset. The results show that, when classifying melanoma, additional pre-training on skin disease does not significantly affect the performance of a CNN which has already been pre-trained on a large number of images from various, less related domains. Additionally, the results indicate that fine-tuning the top convolutional layers of a skin disease pretrained CNN is somewhat beneficial and that some features learned from training on skin disease are less useful for classifying skin cancer.