An automatic skin lesion segmentation system with hybrid FCN-resalexnet

Skin cancer is one of the most common of all types of cancer and has caused the death of many people worldwide and has had a significant increase in the number of populations it has affected in recent years. When diagnosed early, melanom, the most dangerous type, is a cancer type with the highest curability rate. Diagnosis of skin lesions is difficult due to their morphological nature. For this reason, successful results to be achieved in computer-aided automatic diagnosis systems are very important. Previous studies have shown that successful segmentation of skin lesions increases the accuracy achieved by automatic diagnosis systems. Considering these problems, this study on automatic segmentation of skin lesions consists of two stages. In the first stage, the effects of some hyperparameters used in the structure and training of Convolutional Neural Network based deep segmentation architectures on skin lesion segmentation performance were evaluated. In the second phase, a hybrid Fully Convolutional Network based segmentation architecture defined as FCN-ResAlexNet was designed for skin lesion segmentation. When the results were evaluated, the proposed architecture performed better than the most popular segmentation architectures in the literature. It was among the pioneering studies with the performances obtained when compared with other studies in the literature.