Fig 1.
Schematic diagram of the five categories of the positional relationship between the subject and the X-ray detector.
As shown in the figure, the video field of view only includes the chest radiography X-ray detector panel and its immediate vicinity, and is categorized into five classes based on its status(from left to right, the detector states are: “No one”, “Others”,”Too High”,”Too Low” and “Suitable”).
Fig 2.
Image Preprocessing and Data Allocation Flowchart.
Fig 3.
Training Process for Two Types of Convolutional Neural Networks and AI Models.
Fig 4.
Grad-CAM heatmaps generated for each test example in Model A (fold 4) and Model B (fold 1) compared with the original images (from top to bottom, each row corresponds to the detector states: “Too High”,”Too Low”, “Suitable”,”No one” and “Others”).
Table 1.
Accuracy and loss function values for the validation set corresponding to each model fold.
Table 2.
Derivative metrics from confusion matrices generated from the test set for each model fold.
Table 3.
Derivative metrics from confusion matrices and McNemar’s test results for each category test in Model A (fold 4)and Model B (fold 1).
Fig 5.
Confusion matrix visualizations generated for each category test in Model A (fold 4)and Model B (fold 1).
Fig 6.
ROC curves generated for each category test in Model A (fold 4)and Model B (fold 1).
Fig 7.
Calibration curves and Brier scores generated for each category test in Model A (fold 4)and Model B (fold 1).