Chest X-Ray Dataset and Deep Models for Clinical COVID-19 Severity Stratification

The lack of annotated chest X-ray (CXR) datasets reflecting COVID-19 lung damage severity poses challenges for automated clinical diagnosis. This study aims to develop a highquality multi-class CXR dataset with four defined levels of pulmonary severity: Severe, Moderate, Mild and Normal. The dataset was constructed by combining and preprocessing Brixia and COVIDGR datasets, with quality control, class balancing, and lung region segmentation using YOLOv8. To evaluate the dataset, transfer learning was applied using six State-of-the-art models. All models were trained on standardized input with consistent hyperparameters. Among them, Xception achieved the best performance with 97.91 % Accuracy and 99.91 % AUC. Confusion matrix analysis showed high reliability, especially in separating Normal and Severe cases. The results affirm the reliability of the proposed dataset as well as the effectiveness and potential of deep learning models in supporting clinical assessment of COVID-19-related lung injury.