Neural network classification of dermatoscopic images of pigmented skin lesions

Abstract

Today, skin cancer can be regarded as one of the leading causes of death in humans. Skin cancer is the most common type of malignant neoplasm in the body. Rapid and highly accurate diagnosis of malignant skin lesions can reduce the risk of mortality in patients. The paper proposes a neural network classification system of pigmented skin lesions according to 10 diagnostically significant categories. Modeling was carried out using the MATLAB R2020b software package on clinical dermatoscopic images from the international open archive ISIC Melanoma Project. The main convolutional neural network architectures used were SqueezeNet, AlexNet, GoogLeNet, and ResNet101, pre-trained on the ImageNet set of natural images. The highest accuracy rate was achieved using the AlexNet convolutional neural network architecture and amounted to 80.15%. The use of the proposed neural network system for the recognition and classification of dermatoscopic images of pigmented lesions by specialists will improve the accuracy and efficiency of the analysis compared to the methods of visual diagnostics. Timely diagnosis will allow starting treatment at an earlier stage of the disease, which directly affects the percentage of survival and recovery of patients.