The aim of this study was to analyze the possibility of using a mathematical model of logistic regression for the recognition of malignant neoplasms on digital images of the skin. The study used a database containing 6594 digital skin images. At the first stage of the study, digital skin images were segmented to isolate the object under study, in which morphometric and color characteristics corresponding to the parameters of the classical ABCD system were determined. At the second stage, the characteristics were used in the classification into malignant and benign neoplasms using logistic regression. When classifying images, the highest value of the accuracy indicator (67.9 [66.9; 68.8]%) was obtained during classification using logistic regression, built on the basis of the reverse Wald stepwise method. Thus, the logistic regression built on the basis of the reverse stepwise Wald method can be applied in the classification of malignant neoplasms on digital images of the skin, but further research and determination of the optimal parameters are required.

Keywords: mathematical model, digital skin images, logistic regression, image classification, skin cancer

The purpose of this research is analysis of the possibility of using classical neural networks for the malignant neoplasms recognition on skin digital images. In this study was used International Skin Imaging Collaboration database including 6594 dermatoscopic images. At the first stage of the study, digital skin images were classified into malignant and benign neoplasms using the IBM SPSS Statistic tool to automatically select the architecture for artificial neural network. At the second stage was built architecture of artificial neural network, which include one hidden layer. At the third stage was used architecture of an artificial neural network with two hidden layers. In the course of the study digital skin images classified with the highest value of the accuracy indicator (0,752 [0,736 ; 0,768]) during classification using the architecture of an artificial neural network, which includes two hidden layers. The sigmoid was used as the activation function for the hidden layers. The hyperbolic tangent was used as activation function for output layer. With this value of the accuracy, the specificity of the diagnostic method was obtained – 0,813 [0,802 ; 0,824], as well as the value of the sensitivity – 0,665 [0,637 ; 0,691]. Thus, artificial neural networks can be used as a method for skin malignant neoplasms diagnostic on digital images.

Keywords: artificial neural networks, digital skin images, machine learning, image classification, skin cancer