A Hybrid Feature Extraction Network for COVID-19 Classification Based on Chest X-Ray Images Using LGBM

The COVID-19 outbreak stands as one of the most significant global health crises in recent memory, claiming millions of lives worldwide. The urgent and precise identification of the disease holds paramount importance for effective treatment and preventive measures due to its rapid spread. Moreover, distinguishing between COVID-19 and other pneumonia ailments is crucial for accurate diagnosis and subsequent treatment, given their similar symptomatology and physiological impacts. Leveraging x-ray imaging in medical diagnostics offers a rapid and cost-effective means of detection. In this study, a dual-stage feature selection methodology is employed on hybrid or combined deep convolutional characteristics or features extracted from x-ray images. Three convolutional neural networks (CNNs) - Inception V3,ResNet50, and Xception - are utilized in a transfer learning framework. The performance of combining features from multiple CNNs is meticulously analyzed. Initially, features from various models are selected at a specified percentage in the first level of feature selection, followed by a second level to identify the most informative features. Both levels employ the Light Gradient Boosting Machine (LGBM) algorithm for selection. The resultant feature set is then utilized to classify Viral Pneumonia, COVID-19 and non-Covid-19 chest x-ray images using a Support Vector Machines (SVM) classifier. The proposed model's efficacy is evaluated using an open-access dataset, achieving a peak accuracy of 98.49% by leveraging only 35% of the features extracted from the proposed hybrid model consisting of ResNet50, InceptionV3,and Xception networks.