The purpose of the research. The aim of the study is Russian and foreign studies have shown that malignant breast tumors have significantly different impedance from normal tissues. However, bioimpedance analysis has limitations in resolution, as well as in the imperfection of bioimpedance models required to generate input vectors for machine learning systems.Methods. The presented study proposes a multimodal classifier, the raw data for which are obtained through an electrode matrix. It also has three channels for processing the results of bioimpedance analysis, with subsequent aggregation of their solutions. An impedance model of biomaterial is proposed, which allows forming descriptors for medical risk classifiers.Results. Hardware and software for bioimpedance studies have been developed, which include a data collection device for bioimpedance spectroscopy based on an electrode matrix, a device for communication with the object of study, and a device for bioimpedance spectroscopy using an electrode matrix. The software includes interface windows for setting up the bioimpedance research program and training and testing fully connected neural networks. An experimental study of the multimodal classifier on a physical model was conducted using inclusions of higher conductivity (tumor imitation) of various types and sizes in the conductivity range from 1.1 to 1.9 of the background. Based on the obtained images in the two-level neural network of the first channel, the integral risk of breast cancer was determined for all pixels of the image. Statistical studies (ROC analysis) showed sufficient sensitivity and specificity for the screening method - > 0.75. Conclusion. Thus, a new model of intelligent support for medical decision-making has been created, integrating the capabilities of bioimpedance spectroscopy, convolutional neural networks and expert assessment of images generated by bioimpedance mapping. However, current data confirming the possibility of separating benign and malignant breast tumors using bioimpedancemetry methods are very limited, which requires further research in this direction.
Read full abstract