Fig 1.
Overall workflow and algorithm output.
Fig 2.
Model architecture and training workflow.
Fig 3.
An example of the algorithm interface.
The upper toolbar contains basic image viewing functions such as zooming, window adjustment, panning, inversion, rotation, and measurement tools. The left panel displays the examination series, while the central window shows the chest image for AI analysis and visualization. The highlighted region in yellow in the left upper lung field demonstrates patchy, nodular, and linear high-density opacities with partially blurred margins. The right panel presents the AI diagnostic report, including three major sections: (1) AI lesion annotation: displays the detected lesion location and imaging findings (e.g., patchy, nodular, and linear high-density opacities in the left upper lung field). (2) Intelligent auxiliary screening: lists common thoracic abnormalities identified by the algorithm; the red indicators highlight positive findings for infectious lesions, active pulmonary TB, and inactive pulmonary TB. (3) Auxiliary diagnostic report: provides the textual interpretation and summary, including imaging description (Imaging findings) and overall assessment (Imaging evaluation).
Fig 4.
The images collecting process and the evaluation workflow.
Fig 5.
ROC curve of the TB Screening Al Algorithm (CXR).
The values in parentheses shown on the ROC curves represent the x- and y-coordinates of the selected operating points.
Table 1.
Accuracy metrics of the TB Screening AI Algorithm (CXR). Threshold = 0.350 represents the best-performing threshold identified in the previous version of the algorithm, while threshold = 0.085 was determined as the optimal cutoff in the current study.
Table 2.
Accuracy by diagnostic outcomes.
Fig 6.
ROC curve of the TB Screening Al Algorithm (localizer images of chest CT).
The values in parentheses shown on the ROC curves represent the x- and y-coordinates of the selected operating points.
Table 3.
Accuracy metrics of the TB Screening AI Algorithm (Localizer Images of Chest CT). Threshold = 0.350 represents the best-performing threshold identified in the previous version of the algorithm, while threshold = 0.111 was determined as the optimal cutoff in the current study.
Fig 7.
Scatter plot of Al performance in CXR and localizer images of the chest CT in the 105 cases.
Scatterplot showing the correlation between CXR-based and CT-localizer–based AI prediction scores for 105 tuberculosis cases. Each point represents the AI-inferred probability (0–1) of active TB for each patient. A significant correlation was observed (Spearman r = 0.577, P < 0.001).
Fig 8.
ROC curves of CXR and localizer images of the chest CT in the 105 cases.
The values in parentheses shown on the ROC curves represent the x- and y-coordinates of the selected operating points.
Table 4.
Accuracy metrics of the TB Screening AI Algorithm in the 105 Cases (CXR). Threshold = 0.350 represents the best-performing threshold identified in the previous version of the algorithm, while threshold = 0.027 was determined as the optimal cutoff in the current study.
Table 5.
Accuracy metrics of the TB Screening AI Algorithm in the 105 Cases (Localizer Images of Chest CT). Threshold = 0.350 represents the best-performing threshold identified in the previous version of the algorithm, while threshold = 0.041 was determined as the optimal cutoff in the current study.