Enhancing COVID-19 severity assessment with artificial intelligence-based bone suppression technique in chest radiography

Chest radiography (CXR) is widely used for initial respiratory assessment, but its lesion detection capability is typically inferior to that of computed tomography. Several studies have reported that artificial intelligence (AI)-based bone suppression techniques can enhance the accuracy of lesion detection and disease classification. Previously, we developed an AI-based bone suppression system based on dual-energy subtraction principles. However, the subtraction process limited its versatility and introduced significant artifacts. To overcome these challenges, we improved the system to generate bone-suppressed images directly, eliminating the need for subtraction. This study demonstrates the utility of the updated bone suppression system as a pre-processing tool for regression analyses in assessing coronavirus disease 2019 severity. Four regression models – DenseNet, ResNet18, ResNet50, and RegNetY-120 – were employed to predict the severity based on scores annotated by radiologists. Except for DenseNet, all models showed statistically significant improvements in Pearson correlation coefficients (PCCs) when using bone-suppressed images generated by the updated model. The highest PCC, 0.895, was achieved by the ResNet18 model. The direct image generation process improved the clinical practicality of the bone suppression system while reducing artifacts. Furthermore, the significant improvement in linearity suggests that AI-driven bone suppression enhances the visibility of abnormalities and improves the accuracy for pulmonary condition assessments. These advancements could expand the application of bone suppression techniques in various regression analyses, including disease severity, progression, and recurrence risk. Nonetheless, further validation using larger and more diverse datasets, as well as a broader range of prediction models, is necessary.