“…We also evaluated a number of recent techniques, focusing on the U-net architecture (RONNEBERGER et al, 2015), which is widely used for biomedical and other small, specific datasets. Our results reinforce the findings of Codella et al (2017) that adding fully-connected layers to the U-net network (which in its original form, only contains convolutional layers) results in better performance. We also found that using a loss based on the Jaccard metric leads to better results, a similar finding to that of in the same ISIC 2017 Challenge.…”