Deep Learning Based Melanoma Diagnosis Identification

Abstract Malignant melanoma is considered to be one of the deadliest types of skin cancer, and it is responsible for the death of a large number of people worldwide. However, distinguishing whether melanoma is benign or malignant has been a challenging task. Many Computer Aided Diagnosis and Detection Systems have been developed in the past for this task. This paper presents a deep learning framework based approach for melanoma diagnosis and recognition. In the proposed method, the original skin mirror image is first preprocessed and then passed to the VGG16 convolutional neural network for tumor property classification. VGG16 uses smaller convolutional kernels instead of a larger convolutional kernel to achieve a reduction in network parameters and thus improve network performance. The system is trained using segmented RGB images generated from ground truth images of the ISIC2016 dataset, and finally a softmax classifier is used for pixel-level classification of melanoma lesions. In this study, a new method to become a lesion classifier was designed to classify melanoma lesion regions into benign and malignant tumors based on the results of pixel-level classification, and experiments were conducted on two well-established public test datasets, ISIC2016 and ISIC2017, with a final accuracy of 96.1%. The results indicate that convolutional neural networks are suitable for melanoma diagnosis identification. This study is of great relevance for advanced cancer caused by malignant melanoma.