Segmentation of Benign and Malign lesions on skin images using U-Net

Abstract

One of the types of cancer that requires early diagnosis is skin cancer. Melanoma is a deadly type of skin cancer. Computer-aided systems can detect the findings in medical examinations that human perception cannot recognize, and these findings can help the clinicans to make an early diagnosis. Therefore, the need for computer aided systems has increased. In this study, a deep learning-based method that segments melanoma with color images taken from dermoscopy devices is proposed. For this method, ISIC 2017 (International Skin Image Collaboration) database is used. It contains 1403 training and 597 test data. The method is based on preprocessing and U-Net architecture. Gaussian and Difference of Gaussian (DoG) filters are used in the preprocessing stage. It is aimed to make skin images more convenient before U-Net. As a result of the segmentation performed with these data, the education success rate reached 96-95%. A high similarity coefficient obtained. On the other hand, as a result of the training of the preprocessed data, accuracy rate has reached 86-85%.