MOSCOW, December 7. Scientists North Caucasus Federal Universityhave developed a neural network system for recognizing skin cancer, which will reduce the number of false diagnoses. According to the authors, the use of their development as an auxiliary diagnostic method will reduce the influence of the human factor when making medical decisions and increase the accuracy of disease detection. The results of the study were published in IEEE Access.
Skin cancer is one of the most common types of cancer. Highly accurate diagnostics increases the chances of recovery for patients. Therefore, specialists are interested in developing automated auxiliary diagnostic systems.
Scientists from NCFU have developed a multimodal neural network system for classifying oncological skin lesions, sensitive to unbalanced dermatological data.
The proposed system, according to one of the authors of the article, junior researcher at the research laboratory of the Department of Mathematical Modeling of NKFU Ulyana Lyakhova, allows to reduce the number of false negative predictions through the use of a modified cross-entropy loss function and analysis of heterogeneous dermatological data with the stage of preliminary cleaning of hair structures.< br />As stated in NCFU, the recognition accuracy in ten diagnostic categories for the proposed intelligent system was 85.2 percent. In addition, the system recognizes pigmented skin lesions 15 percentage points more accurately than the visual diagnosis of medical practitioners.
The system developed in Russia allows achieving accuracy higher than that of similar foreign systems from Germany, Austria, China, scientists noted.
>
“The use of the developed complex by dermatologists as an auxiliary diagnostic method will reduce the influence of the human factor when making medical decisions, significantly reduce the number of false diagnoses and increase the accuracy of early detection of skin cancer,” said Pavel Lyakhov, head of the department of mathematical modeling of NKFU.
1 of 2
2 of 2
1 of 2
2 of 2
As noted by co-author of the study, Associate Professor of the Department of Mathematical Modeling at SFU, Diana Kalita, the system, using various types of neural networks, simultaneously analyzes the image of moles and basic patient data (gender, age, localization of pigmented neoplasm). But before analysis begins, images of pigment spots are processed using certain filters that make it possible to remove details that interfere with a more accurate classification (for example, hair and other noise effects).
«The non-obvious relationship between the processed data and diagnostic results is extracted using due to additional neural network study of information between modalities. Thus, neural networks are able to use additional data by integrating several modalities into a common structure,» explained Ulyana Lyakhova.
In the future, the NCFU research team plans to build more complex ensemble systems for neural network analysis of dermatological data.
The research was supported by the Russian Science Foundation.