Detection of Melanoma using Deep Learning Techniques

Melanoma is a type of skin cancer, which, if detected in its early stages, is highly curable. But its detection is hard, even under expert supervision. This paper is an attempt to make detection of Melanoma using deep learning techniques more efficient and reliable compared to existing techniques. The overall approach followed is to build a two-stage network. The first-stage network targets accurate segmentation of the skin lesion, from the actual dermoscopic images. The second-stage network is a classification network to predict the presence of Melanoma in the sample. For the segmentation stage network, both the U-NET and FCRN methods were implemented. For the classification network, the DRN architecture was implemented. In order to enhance the achieved results, the step-decay technique to modify learning rates was used. Using both binary cross-entropy and weighted binary cross-entropy improved the achieved results, driving towards better accuracy of detection.