Using Machine Learning Methods to Analyze Microcirculation

Abstract

Analysis of microcirculation images can reveal early signs of microvascular dysfunction, which is caused by a synergistic effect from numerous acute and chronic diseases. Quantification of capillary density and distribution of capillaries on microcirculation images acts as a biological marker. Thus, the reduced functional density of capillaries contributes to pathological tissue hypoxia in sepsis and diabetes; abnormal vascular permeability contributes to diabetic nephropathy and stroke; reduced arteriolar blood flow and impaired ability to regulate vascular tone are involved in the pathogenesis of hypertension; flow-mediated vasodilation is impaired in patients with coronary artery disease; arteriolar networks show abnormal morphology and distribution of blood flow in metabolic syndrome; regional perfusion defects occur in the cerebral microcirculation after cardiac arrest. Quantification of the above described biological markers is labor intensive, time consuming and subject to interobserver variability. In view of the identified problem, the scientific community involved in the study of microvascular systems is focused on standardizing the analysis of microcirculation images using automated methods. Since at the moment microvascular analysis is carried out mainly by a manual method, it is necessary to digitize the knowledge of experts. We are at an important stage in the transfer of knowledge from experts to software and computer systems using machine learning methods. In this paper, we have chosen a methodology for using expert knowledge that compensates for the sparseness of the data. Using it in training allows you to get forecasting models with good predictive power. At the first stage, the expert checks the selected, especially informative charts of this model. In the second step, the expert indicates how the graphs of gray activity confirm or contradict his expectations of the shape and activity of the capillary network. And in the third and final step, the expert knowledge of the form thus defined is incorporated into a new prediction model that strictly complies with all imposed form constraints.

Speaker

Reschetnikov Vladislav
ITMO
Russia

Discussion

Ask question