Machine Learning Based on Computed Tomography Pulmonary Angiography in Evaluating Pulmonary Artery Pressure in Patients with Pulmonary Hypertension

Figures

• 

  Figure 1. The overall workflow. After the selection and exclusion of patient population and CTPA images acquisition, eight substructures of heart and pulmonary arteries were automatically segmented as regions of interest (ROIs). Morphological features and their second-order features are extracted from these ROIs. Then, the essential features were selected through statistical methods and then reconstructed to reduce the redundancy of the features matrix. In the last step, meaningful features and corresponding results are set as an input through machine learning algorithms to obtain a regression model and a classification model.

• 

  Figure 2. Segmentation framework architecture for accurate segmentation of the substructures of pulmonary artery and heart, respectively. The proposed segmentation framework consists of full-resolution 3D U-Net; contours feature extraction; and positive mask decision.

• 

  Figure 3. Some of the morphological features on CTPA are shown above. (a) Diameter of main pulmonary artery (MPAd), (b) diameter of ascending aorta (AAd), (c) diameter of right pulmonary artery (RPAd), (d) area of LPA (blue), RPA (green), MPA (red), and AA (yellow), (e) diameter of left ventricle, (f) diameter of right atrium, (g) diameter of right ventricle, (h) area of LV (red), RV (green), LA (blue), and RA (yellow). They were all assessed on transversal images in the diastole. The features extracted from the original CTPA images were first organized in a table, sorted by the Spearman’s rank correlation coefficient between one feature and mean pulmonary artery pressure. Since any two features we selected may contain a strong correlation, we removed redundant features by calculating the Pearson coefficient of all features with each other. Then we selected r features from all features we obtained. The considerately selected features were then normalized and feature dimension reduction was performed by principal component analysis. Finally, we chose m features for the regression and classification model.

• 

  Figure 4. The performance of the proposed network framework. (a,d) are the original image of heart and pulmonary artery, respectively; (b,e) are the segmentation outputs of nnU-Net; (c,f) are the segmentation outputs of the proposed network framework.

• 

  Figure 5. Bland–Altman analyses for features assessed by AI automatic and manual measurements show that the metrics measured by our automatic measurement method are in accordance with the ground-truth measured manually by experienced physicians.

• 

  Figure 6. (a) CatBoost regression on mPAP (MSE = 28.37), (b) SVM regression on mPAP (MSE = 60.24), (c) XGBoost regression on mPAP (MSE = 12.82), (d) XGBoost regression on dPAP (MSE = 16.94), (e) XGBoost regression on sPAP (21.88), (f) XGBoost regression on TPR (MSE = 69,862.59).

• 

  Figure 7. Area under the receiver operating characteristic curve (AUC) shows XGBoost model performs better than CatBoost and SVM model. The AUC of XGBoost model classification on mPAP and sPAP is 0.911 (p < 0.001) and 0.833 (p = 0.0057), respectively.