A.A. Ignatiev

The importance of melanoma early diagnostics is a key factor in improving patient survival. Prevention and early detection of melanoma have led to progress in skin cancer diagnostics, including the technologies development in the sphere of artificial intelligence (AI). Assessment of possibilities and limitations of AI in dermatology has a prime importance for its future development and practical application. AI, realized as a chat-bot format, can help in melanoma early diagnostics and has a special value at the interface between primary and secondary health care. The new technology is an excellent opportunity for doctors to learn melanoma diagnostics in a playful way, realizing the capabilities of this new diagnostic tool and promoting its safe and ethical integration into the health system. As a part of this solution development, it’s planned to expand the educational data set to check the dermatologists’ skills of melanoma identification by dermatoscopic images. In addition, it’s planned to add unclassified images to classify them in future by system users, who received the best results in statistics of proper answers. The introduction of new images by doctors will allow to continue the learning of AI algorithm to improve disease’s machine detection based on uploaded image.

The approach to diagnosis and treatment in dermatovenerology has changed significantly with the introduction of new technologies and inventions. Computer algorithms have been used to assist dermatovenerologists in diagnosing diseases, including malignant melanoma of the skin. Instrumental methods for diagnosing skin melanoma, based on the processing of dermatoscopic images, can be improved by using artificial intelligence. Machine learning is a subset of artificial intelligence in which computer programs automatically learn from experience without explicit operation instructions. The most promising direction is deep learning of artificial neural networks. In this paper, several models of artificial neural networks were studied to assess the prospects for their use for the early diagnosis of skin melanoma. The efficacy of the various models was analyzed using the HAM10000 dataset image classification. Models of convolutional neural networks, pre-trained neural networks with transfer learning, and the “soft attention” mechanism were evaluated: Xception, ResNet50, RAN50, SEnet50, ARL-CNN50, Inception Resnet V2 — IRv2 12x12+SA, IRv2 5x5+SA. The results showed that the most promising model was Inception Resnet V2 with the addition of a «soft attention» mechanism.