Результати застосування методів машинного навчання у дослідженні COVID-19 асоційованої кардіопульмональної патології за даними комп'ютерної томографії

Background. The COVID-19 pandemic, caused by SARS-CoV-2, has significantly impacted global health, emphasizing the importance of efficient diagnostic methods. Computed tomography (CT) imaging plays a crucial role in identifying COVID-19-associated lung pathologies, yet manual analysis of extensive imaging data remains burdensome. Machine learning (ML) methods offer promising automated solutions to expedite diagnostics and reduce workload on radiologists. Objective. To evaluate the effectiveness of machine learning algorithms, specifically convolutional neural networks (CNN) and texture analysis methods, in automated detection and classification of COVID-19-related cardiopulmonary pathology using chest CT imaging. Methods. This study analyzed chest CT datasets obtained from clinical resources, which included images from patients with COVID-19 exhibiting ground-glass opacities, crazy-paving patterns, and consolidations. Regions of interest (ROIs) were segmented and classified using various machine learning approaches: CNN combined with gray-level co-occurrence matrix (GLCM) texture analysis, logistic self-organizing forest (LSOF), group method of data handling (GMDH), and ensemble methods including random forest, XGBoost, LightGBM, and random forest of optimal complexity trees (RFOCT). Results. The CNN and texture-based hybrid classifiers achieved high accuracy, with overall classification accuracies ranging from 83% to 99%. Specifically, ground-glass opacity identification reached up to 100% accuracy, while crazy-paving patterns and consolidations showed slightly lower accuracies (71–95%). The ensemble method RFOCT achieved the highest accuracy (89%) in differentiating acute COVID-19 from Long COVID. Additionally, methods incorporating texture analysis significantly enhanced the accuracy and informativeness of CT-based diagnostics. Conclusions. Machine learning algorithms, particularly CNNs and advanced texture analysis, demonstrate significant potential in automating the diagnosis of COVID-19-associated lung pathology. These approaches not only increase diagnostic efficiency but also facilitate the detection of subtle pathological changes, crucial for clinical decision-making and patient management during epidemiological crises. Future research should address current limitations related to dataset size, computational complexity, and generalizability to further enhance clinical applicability.