Fig 1.
Workflow diagram illustrating the overall experimental design.
It described the flow of CXR images through labeling process followed by training transfer learning models using multicenter data and evaluating the models with internal multicenter testing data and independent external testing data. The labeled images were chosen from the initial total dataset according to diagnostic criteria and sufficient quality. CM, confusion matrix; MCC, Matthews correlation coefficient; ROC, receiver operating characteristic; AUC, the area under the curve; MAE, mean absolute error.
Fig 2.
Schematic of labeling and training process.
A Patient with or without PH receiving CXR screening (a). The same patient receiving TTE to identify PH by measuring PASP value with cutoff of 40 mmHg (b). Using retrospective CXR and corresponding PASP tags to train transfer learning models that can be deployed to diagnose PH and predict PASP according to the CXR of new patients (c). CXR, chest X-ray; TTE, transthoracic echocardiography; PH, pulmonary hypertension; PASP, pulmonary artery systolic pressure.
Fig 3.
The AUC/ROCs for detection of PH from normal.
Comparision of AUC/ROCs in internal test. The InceptionV3 model had an AUC (0.970) greater than the other two models (a). Comparision of AUC/ROCs in external test. The InceptionV3 model provided the best AUC (0.967) compared to the other ones (b). The AUC performed by InceptionV3 model in external test was slightly lower than that in internal test (0.967 VS 0.970 respetively) (c). ROC, receiver operating characteristic; AUC, area under the ROC curve; PH, pulmonary hypertension.
Table 1.
Comparison of performances of deep learning and manual in internal and external test.
Fig 4.
Prediction of exact PASP value by best model (InceptionV3) based on chest radiographs.
Prediction of exact PASP value by InceptionV3 model in internal test, with a MAE of 7.45 (a). Prediction of exact PASP value by InceptionV3 model in external test, and the MAE was 9.95 (b). PASP, pulmonary artery systolic pressure; MAE, mean absolute error.
Table 2.
Comparison of MAEs for regression prediction of PASP in internal and external test.
Fig 5.
Comparision of the advantage and defect of different ways to screen spicious PH patients.
PH, pulmonary hypertension; RHC, right heart catheterization; TTE, transthoracic echocardiography; CT, computed tomography.
Fig 6.
Visualization of CNN models using Grad-CAM to classify PH from normal based on radiograph images.
The highlighted areas indicatied by red arrows are discriminative features for identification of PH. The Grad-CAM of patient without PH (a). The Grad-CAM of patient with PH (with PASP of 56mmHg) (b). CNN, Convolutional Neural Network; PH, pulmonary hypertension; Grad-CAM, Gradient Class Activation Map; PASP, pulmonary artery systolic pressure.