Lesion detection on skin images using improved U-net

Abstract

One of the most prevalent cancers in humans is skin cancer. The deadliest form of skin cancer is malignant melanoma and the incidence rate has increased rapidly in recent years. In the treatment of melanoma, early diagnosis is very critical. It is difficult and time consuming to automatically detect melanoma from images taken from dermoscopy devices. Computer-aided systems are needed, therefore. In this paper, a deep learning-based method for melanoma segmentation and classification with color images taken from dermoscopy devices is proposed. This technique uses ISIC 2017 International Skin Imaging Collaboration. In this paper, for segmentation and classification measures, 1317 skin images taken from the ISIC archive were used. The approach is based on the architecture of Preprocessing, U-Net and VGGNet. Operations such as mean subtraction, image normalization, image cropping, and scaling are implemented in the preprocessing phase. It is intended to make pictures of the skin more convenient before segmentation. The training precision rate and jaccard similarity coefficient reached 93% as a result of segmentation with these results, and the dice coefficient reached 79%. The accuracy rate is 85.5% as a result of the classification in the two-class dataset in the pre-trained VGG16 network. The accuracy rate of dataset classification obtained with cross-validation is 95.86%.