Сonvolutional neural networks in the diagnosis of skin neoplasms

Abstract

The problem of using artificial intelligence technologies in the diagnosis of skin neoplasms is considered. Dermatoscopic images for 8 nosologies were considered as the object of research. The melanoma was one of them. Melanoma is responsible for the most deaths of all skin cancers. The aim of the study was to evaluate the effectiveness of the use of pre-trained convolutional neural networks for the classification of skin neoplasms. A classification algorithm for an ensemble of convolutional networks was proposed. Pre-trained neural networks have been studied to form the ensemble. Neural network samples were selected from a set of neural networks that have proven themselves in the ImageNet Large Scale Visual Recognition Challenge. According to the results of the experiment the best three of the eight convolutional neural networks were selected for inclusion in the ensemble – MobileNet_v2, ResNet_152, ResNeXt_101_32x8d. The experiment was conducted on a sample of 10015 images representing 8 nosologies. The average classification accuracy for all nosologies was 79%. The paper highlights the features of ensuring information security when using telemedicine diagnostic technologies using the proposed approach in the recognition of images of skin neoplasms. The results of the work can be used in the design of medical decision support systems for the diagnosis of malignant skin neoplasms (including melanoma).

Keywords

• artificial intelligence, convolutional neural networks, image recognition, melanoma diagnostics, telemedicine technologies, information security.