| Skin cancer is one of the most diagnosed cancers according to the World Health Organization and one of the most malignant. Unfortunately, still the available annotated data are in most cases not enough to successfully train deep learning algorithms that would allow highly accurate predictions. In this paper, we propose the utilization of transfer learning to fine-tune the parameters at the very last layers of a pre-trained a deep learning neural network. We expect that a limited number of skin lesion images is enough to affect significantly the later data-specific layers. Furthermore, we propose a pre-process step for skin lesion images that segments and crops the lesion, whereas smooths the effect of image masking, thus enhancing the network’s classification capabilities. The reported results are very promising, since the overall accuracy, as well as the accuracy of individual class classification improved in 7 out of the 8 classes, suggesting future developments in medical diagnosis through pre-trained deep learning models and specialized image prefiltering. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.