Abstract
Melanoma is fatal skin cancer, whose appearance is quite similar to nevus and diagnostic accuracy by expert dermatologists is estimated to be about 75-84%. With this situation, studies on automated melanoma discrimination have been investigated, however it is difficult to design and extract effective features to build an accurate classifier as well as an appropriate segmentation of tumor area from surrounding skin as a pre-processing. In this study, we applied deep convolutional neural networks (DCNN) for automated melanoma discrimination system. Since DCNN automatically extracts and learns effective image features as a part of its training and performs accurate inference with them, abovementioned troublesome tasks becomes unnecessary. We used a set of 319 dermoscopy images, in which 244 nevi and 75 melanomas and trained our DCNN with distinctive inventions on the training approach. Our system achieved a sensitivity (SE: melanoma detection accuracy) of 82.8% and a specificity (SP: benign detection accuracy) of 90.4% under the 5-fold cross-validation test.
