Comparació de xarxes de segmentació semàntica d'aplicacions dermatològiques en l'entorn de l'ISIC Challenge

Abstract

[en] Cancer affects millions of people every single year around the world. Making it the most common type of cancer, skin cancer owns the record of being responsible to end the life of thousands of human beings. Therefore finding a way to detect it and to classify it is a major issue. At the same time, we cannot forget about the importance of prevention and research about new ways to cure it. Recently, there’s been a significant improvement in the field of neural networks, more specific in the design and research about different kinds of networks that help to classify pixels from an image. This project will analyze multiple sorts of models of neural networks such as U-net or DeepLab v3+, to further get an estimation of its performance regarding semantic segmentation using dermatologic images as a training set. Finally, we will emulate ISIC Challenge 2017 to see how well our networks do. Yielding an accuracy of almost 94% and a F-Score around 86% DeepLab v3+ model has claimed the best result of all of them. As a result, further research in semantic segmentation’s field has been encouraged so melanoma’s detection process can be catalyzed. That way survival rate will directly be increased.