Chest radiography deep radiomics-enabled aortic arch calcification interpretation across different populations

Figures

• 

  Figure 1. The strategy and algorithm for deep learning and transfer learning application to chest radiographs based on the three independent dataset CKD, chronic kidney disease; CXR, chest X-ray; NTUH, National Taiwan University Hospital.

• 

  Figure 2. Confusion matrices for evaluating the 4 deep learning and ensemble model prediction results in general population/older adults (upper row) and in patients with pre-end-stage kidney disease (lower row) (A and F), VGG model; (B and G), Xception model; (C and H), DenseNet model; (D and I), MobileNet model; (E and J), ensemble model.

• 

  Figure 3. ROC curves of different classification models tested in this study Area under the receiver-operating characteristic curves for the 4 deep learning and our ensemble model prediction results in general population/older adults (A) and in patients with pre-end-stage kidney disease (B). AUC, area under the ROC curve; ROC, receiver-operating-characteristic curve.

• 

  Figure 4. Chest radiographs and the respective attention maps Three examples of deep-learning-facilitated attention areas captured by our approach in general population/older adults (A) and in patients with pre-end-stage kidney disease (B). Left, original image; middle, attention maps; right, overlapping image. VC, vascular calcification.

• 

  Figure 5. Examples of attention areas between general population/older adults (left) and patients with pre-end-stage kidney disease (ESKD) (middle), with differentially weighted areas shown by the red shadow (right) (A) attention to the upper mediastinum. (B) Attention to upward aortic branches. (C) Attention to the ascending aorta. Red arrowheads denote areas emphasized particularly in patients with pre-ESKD.

• 

  Figure 6. Examples of false-positive chest radiographic images misclassified by our ensemble model (A) example 1. (B) example 2. (C) example 3. Red arrowheads denote areas potentially implicated in the process of mis-classification.