Validating the effectiveness of an AI algorithm for pulmonary tuberculosis screening using chest X-ray: Retrospective study and test accuracy with localizer images of the chest CT.
China accounted for 6.8% of global TB cases, and most patients are first diagnosed in general hospitals where chest X-rays (CXR) are widely used for early TB detection. To facilitate diagnosis in resource-limited settings, our study evaluates a CNN-based AI model trained on Chinese CXR data (JF CXR-1 v2), including its experimental application to CT localizer images.
This retrospective study was conducted at China-Japan Friendship Hospital, including 290 CXR images and 433 CT localizer images from TB patients diagnosed between 2017 and 2021. The AI algorithm's diagnostic performance was assessed using sensitivity, specificity, accuracy, Kappa value, and AUC from ROC analysis.
The AI algorithm demonstrated high diagnostic performance on CXR images, achieving an AUC of 0.960 with 91.7% sensitivity and 92.7% specificity in bacteriologically confirmed TB cases. On localizer images of the chest CT, while the performance was more modest (AUC 0.719), a significant correlation between CXR and CT predictions in 105 paired cases suggests potential for cross-modality application with further validation.
The algorithm shows decent diagnostic capability for the CXR samples in this study. This AI algorithm developed based on CXR can, to some extent, identify the imaging features of pulmonary TB when applied to localizer images of chest CT.
This retrospective study was conducted at China-Japan Friendship Hospital, including 290 CXR images and 433 CT localizer images from TB patients diagnosed between 2017 and 2021. The AI algorithm's diagnostic performance was assessed using sensitivity, specificity, accuracy, Kappa value, and AUC from ROC analysis.
The AI algorithm demonstrated high diagnostic performance on CXR images, achieving an AUC of 0.960 with 91.7% sensitivity and 92.7% specificity in bacteriologically confirmed TB cases. On localizer images of the chest CT, while the performance was more modest (AUC 0.719), a significant correlation between CXR and CT predictions in 105 paired cases suggests potential for cross-modality application with further validation.
The algorithm shows decent diagnostic capability for the CXR samples in this study. This AI algorithm developed based on CXR can, to some extent, identify the imaging features of pulmonary TB when applied to localizer images of chest CT.