Skin cancer detection using multi-scale deep learning and transfer learning

Skin Cancer is on the rise and Melanoma is the most threatening type among the skin cancers. Early detection of skin cancer is vital in order to prevent the cancer to be spread to other parts. In this paper a transfer-learning based system is proposed for Melanoma lesions detection. In the proposed system first, the images are preprocessed for removing the noise and illumination effect. In the next step a convolutional neural network is trained based on transfer learning using the weights of ImageNet data set. In the third step the network is fine-tuned to become more specialized for detecting the Melanoma versus other types of benign cancers. The proposed system uses the information from the image in 3 stages. In each stage the focus will be more concentrate on the center on the image where the suspicious part is. The results from these parts are combined and applied to a fully connected neural network. Results shows the superiority of the proposed methods compare to other state-of-the arts methods.