Neural Network Skin Cancer Recognition with a Modified Cross-Entropy Loss Function

Abstract

Skin cancer is currently one of the most common types of human cancer. Due to similar morphological manifestations, the diagnosis of malignant neoplasms is difficult even for experienced dermatologists. Artificial intelligence technologies can equal and even surpass the capabilities of an oncologist in terms of the accuracy of visual diagnostics. The available databases of dermoscopic images and statistical data are highly unbalanced about “benign” cases. When training neural network algorithms on unbalanced bases, there is a problem of reducing the accuracy and performance of models due to the prevailing “benign” cases in the samples. One of the possible ways to solve the problem of unbalanced learning is to modify the loss function by introducing different weight coefficients for the recognition classes. The article proposes a neural network system for the recognition of malignant pigmented skin neoplasms, trained using a modified cross-entropy loss function. The accuracy of recognition of malignant neoplasms of the skin in the proposed system was 88.12%. The use of the proposed system by dermatologists-oncologists as an auxiliary diagnostic method will expand the possibilities of early detection of skin cancer and minimize the influence of the human factor.