Image processing of human blood cell aggregates using artificial intelligence in norm and pathology

Abstract

Blood is the most important fluid of the human body, life-sustaining of organs and tissues and performing many different physiological functions. The erythrocyte aggregation (EA) is ability of erythrocytes to form linear or three-dimensional structures consisting of the cells. EA is one of the most important factor affecting blood viscosity, and as a result its circulation.

Aggregation is caused by the presence of various macromolecules in plasma, such as fibrinogen and albumin. The aggregation may also be caused in environment solutions of high molecular weight water-soluble polymers, for example, in dextran [1]. At the same time, the shape of the aggregates depends on the concentration of dextran [2]. In many pathological conditions, such as SARS-CoV-2 virus, lung cancer, cognitive, cardiovascular and age-associated diseases, an increase in the content of high-molecular proteins is observed; therefore, dextran can be used to model specific pathological conditions [3].

Glutaraldehyde is an organic substance that is used in biomedicine as a fixative due to its ability to cross-link proteins. Glutaraldehyde reduces the aggregation ability of red blood cells (RBC). Fixing agents can be used to simulate low RBC deformability in different pathologies, such as malaria, sickle cell anemia, sepsis, diabetes and peripheral vascular diseases [4].

The purpose of this work is to demonstrate using a neural network the possibility of classifying images of human erythrocyte aggregates in stasis after their disaggregation in applied shear stresses in the microfluidic chamber in various environment media: in autologous plasma and in solution with dextran (70 kDa), as well as during incubation of cells with glutaraldehyde. The concentrations of glutaraldehyde and dextran (0.045 mg/ml and 50 mg/ml, respectively) in the erythrocyte suspension were chosen so that it was visually difficult to determine the medium and the effect of the fixative on the cells.

As a result of the work, a dataset with large number (3000) of erythrocyte aggregates images in different environments was obtained, and a convolutional neural network was trained, which predicts the environment in which cells are located from the image of aggregates and detects the effects of glutaraldehyde. The neural network has shown high efficiency (accuracy > 0.9) in detecting a slightly noticeable unnatural state of cells. Using gradient methods, the features used by the neural network for image classification were analyzed [5].

The study was performed with financial support within the framework of the Development Program of the Interdisciplinary Scientific and Educational School of Moscow University "Photon and Quantum Technologies. Digital medicine" (project No. 23-Ш06-03).

Speaker

Ladynin Aleksandr Ivanovich
Lomonosov Moscow State University, 1-2 Leninskie Gory, Moscow, Russian Federation, 119991
Russian Federation

Discussion

Ask question