ИСТИНА

Objective: To evaluate the effectiveness of neural network methods for rapid detection of red blood cells and measurement of geometric parameters in a microscopic image of a blood smear. Materials and methods: The study analyzed photographs from a wet blood smear microscope, which is whole blood diluted in PBS at a concentration of 1:200. Glass cuvettes were prepared, into which the resulting cell solution was poured: the slide was coated with a thin layer of human albumin at a concentration of 10 mg per 1 ml, then a cover glass was attached to double-sided tape with a thickness of about 100 μm. The solution with blood is poured into the cuvette and left undisturbed in a horizontal position for several minutes so that the cells settle to the bottom of the cuvette, forming a sparse monolayer. The task of isolating individual erythrocytes in the image and constructing the contours of these cells using a neural network was considered. To do this, we manually labeled 60 microscope photographs containing approximately 4,500 cells. Based on the resulting database of labeled cells, a YOLOv8 convolutional neural network was trained. The trained model was applied to other images to automatically measure the geometric parameters of red blood cells. Results. As a result of the work, about 2,000 images of wet blood smears from healthy donors and donors with different classes of heart failure were analyzed. Using a neural network, about 200,000 red blood cells were processed in 90 minutes. The neural network was able to detect 90% of red blood cells. The results of measuring the geometric parameters of cells by the neural network are comparable to the data obtained by classical image processing methods. A correlation was also found between the average size of red blood cells and the class of heart failure. Conclusion. Neural network methods allow for rapid and reliable analysis of blood smear images, determining the number and parameters of red blood cells. This approach can be used to automate laboratory diagnostics.