Hybrid Classifiers Based on CNN, LSOF, GMDH in COVID-19 Pneumonic Lesions Types Classification Task

The paper proposes new network tools for solving classification problems based on the hybridization of convolutional networks, self-organized forests, and Group Method of Data Handling classifiers. To determine the types of lung lesions in COVID-19 on computed tomography slices the new solutions have been applied. The features of the texture obtained from the matrices of GLCM, GLRLM, GLSZM, GLDM, NGTDM statistics are used here as classification features. The mechanism of transformation of texture matrices into class-oriented features based on hybrid architectures of neural networks is developed for this. The concept development of self-organized forest according to the principles of GMDH with the use of logistical mechanisms and optimization of voting functions (LSOF) is proposed. Classifiers based on CNN, LSOF, and GMDH are used to determine the type of lung lesions “ground-glass”, “crazy-paving”, “consolidation”on CT images of patients. The classification results obtained by the proposed algorithms have been compared with other modern methods. Data for the research were provided by the State Institution “F.G. Y anovsky National Institute of Phthisiology and Pulmonology of the NAMS of Ukraine”.