Automatic Detection of Skin Cancer Melanoma Using Transfer Learning in Deep Network

As the deadliest type of skin cancer, melanoma has a high mortality rate and takes away thousands of lives in the UK every year. However, if detected at earlier stage, the survival rate largely increases. With the development of machine learning, many well-known pre-trained models were used to detect melanoma accurately through imaging analysis. The overall performance is far beyond skillful human experts. This paper examined the performance of a pre-trained modelVisual Geometry Group network (VGG) on International Skin Imaging Collaboration (ISIC) 2019 challenge dataset in automatically classifying melanoma and non-melanoma diseases. The highest accuracy achieved was 0.9067 with AU ROC over 0.93. Ablation studies illustrated potential factors that could affect model performance, including training data size, frozen layers, classifier nodes and data augmentation methods.