Extraction and evaluation of features of preterm patent ductus arteriosus in chest X-ray images using deep learning

Figures

• 

  Fig. 1. Flowchart of the data used in this study. The data were collected from the Jeonbuk National University Hospital Neonatal Intensive Care Unit from 2014 to 2023 and labeled by a pediatric cardiologist. The labeling criteria to diagnose severe PDA were as follows: a transductal diameter of 1.5 mm or more, or a left atrial diameter:aortic diameter ratio of 1.4:1 or more. Abbreviations: JBUH, Jeonbuk National University Hospital; NICU, Neonatal intensive care unit; CXR, Chest X-ray; sPDA, severe patent ductus arteriosus.

• 

  Fig. 2. Feature extraction and classification pipeline. The global average pooling layer averages the feature maps of the last convolutional layer to generate a feature vector. For example, the 1st element of the feature vector is the average of the 1st feature map.

• 

  Figure 3. Results of the analysis of the sPDA prediction model. a. Top 10 mean absolute SHAP values of the features, calculated on the test dataset. b. Examples of images having the highest and lowest feature_173 values (i.e., 174th element of the feature vector). The images with low feature_173 values exhibit clear lung fields and normal hearts, while the images with high feature_173 values exhibit more opaque lung fields and larger hearts. c. Mean absolute SHAP values of the ratio-based features on the test dataset. d. Beeswarm plot of SHAP values of the features on the test dataset. Negative SHAP values indicate that the feature value impacts the model prediction negatively, while positive SHAP values indicate that the feature value impacts the model prediction positively. The color of the points indicates the feature value. In the case of the gender feature, red indicates female and blue indicates male.

• 

  Fig. 4. Examples of feature maps of the last convolutional layer before the global average pooling layer and GradCAM++ visualizations of the test images. The visualized feature maps correspond to the three feature values that showed the highest SHAP values (feature_173, feature_347, and feature_1011). Label refers to the sPDA label of the test image, and Pred refers to the predicted probability of sPDA, calculated by the CNX-based feature extraction model.

• 

  Fig. 5. Definition and analysis of the hypothesized radiographic features. (a) CXR images highlighting the proposed feature definitions. The upper heart width is defined as the width of the heart measured at the upper 10% of its total height. The widths are measured based on the heart and thorax masks generated by the segmentation models. (b) Histograms of the features in sPDA-positive and sPDA-negative groups. For accurate comparison, only one image was randomly selected from each patient. The p-values were calculated from z-tests.

• 

  Fig. 6. Waterfall plots of the ratio-based sPDA prediction model. The plots visualize the impact of each feature on the model’s prediction for each test image. The bottom of the plot represents the expected value of the model prediction, and each row represents how the features contribute to the prediction for each test image. Red indicates positive and blue indicates negative contribution to the output of the model. The x-axis represents the log-odds, so negative values indicate a probability of sPDA less than 0.5, and positive values indicate a probability of sPDA greater than 0.5. Abbreviations: CTR, cardiothoracic ratio; Gender: 0 is female, 1 is male; BW, Birth weight (g); GA, Gestational age (weeks); UHW, Upper heart width; MHW, Maximum heart width; TW, Thorax width.