Chest Computed Tomography-Based Radiomics for the Diagnosis and Prognosis of Pulmonary Hypertension

Abstract

Background: Imaging technique has emerged as an innovative tool for diagnosing and monitoring patients with pulmonary hypertension (PH). Current studies in radiomics primarily focus on hemodynamics and cardiac function, whereas the assessment of pulmonary vessels is often neglected. This study aims to investigate the diagnostic and prognostic value of computed tomography pulmonary vascular radiomics in PH.

Methods: This multicenter study enrolled 193 patients with PH and 193 controls (102 symptomatic non-PH cases and 91 healthy volunteers) for diagnostic analysis, with external validation in 38 patients with PH and 38 controls. For prognostic analysis, 166 patients with PH were prospectively followed (median follow-up: 16 months; 96 clinical deterioration events), with external validation in 32 patients with PH (median follow-up: 7 months; 11 events). Pulmonary vascular radiomics features extracted from chest computed tomography were used to develop predictive models for PH diagnosis and prognosis.

Results: The diagnostic model integrating 7 radiomic and 3 clinical features achieved an area under the curve of 0.984 (95% CI, 0.959-0.995) in the derivation cohort and 0.980 (0.901-1.000) in external validation. For exploratory pulmonary arterial hypertension subtyping, the model incorporating 8 radiomic and 2 clinical features yielded an area under the curve of 0.898 (0.825-0.972) internally and 0.877 (0.352-1.000) externally. The prognostic radiomic-clinical model outperformed the European Society of Cardiology 4-strata risk assessment, with a 2-year area under the curve of 0.866 (0.8-0.942) versus 0.709 (0.648-0.789).

Conclusions: The radiomics-based models have strong diagnostic and prognostic capabilities for PH and can also successfully differentiate pulmonary arterial hypertension. This suggests the potential of radiomics to discern PH with different clinical risks, which may facilitate personalized drug therapy.

Keywords: computed tomography; diagnosis; machine learning; prognosis; pulmonary hypertension; radiomics.

Figure 1. Study flow chart.

AUC indicates area under the curve; CT, computed tomography; LASSO, least absolute shrinkage and selection operator; LRC, logistic regression classifier; NPAH Non‐pulmonary arterial hypertension (encompassing WHO Groups 2–5); NPH, nonpulmonary hypertension; PAH, pulmonary arterial hypertension; PH, pulmonary hypertension; RFC, rando forest classifier; RHC, right heart catheterization; ROC, receiver operating characteristics; and SVC, support vector machines classifier.

Figure 2. Segmentation of pulmonary parenchyma and pulmonary vessels from chest CT.

A, Masks and 3D reconstruction of the cross‐section, the coronal plane, and the sagittal plane of pulmonary parenchyma. B, Masks and 3D reconstruction of the cross‐section, the coronal plane, and the sagittal plane of pulmonary vessels. 3D indicates 3‐dimensional; and CT, computed tomography.

Figure 3. Model performance of radiomic models and joint models in diagnosing PH.

A, ROC curve of 6 models in the derivation cohort. B, Calibration curve of 6 models. C, ROC curve of 6 constructed models in validation cohort. D, Decision analysis graph of LRC_TRV, LRC_PA/AO, and LRC_JM model. AUC indicates area under the curve; LRC_JM indicates joint model established by logistic regression classifier; LRC_PA/AO, computed tomography‐measured diameter ratio of the main pulmonary artery and the aorta diagnostic model established by logistic regression classifier; LRC_RM, radiomic model established by logistic regression classifier; LRC_TRV, echocardiographic measured tricuspid regurgitation velocity diagnostic model established by logistic regression classifier; RFC_JM, joint model established by random forests classifier; RFC_RM, radiomic model established by random forests classifier; ROC, receiver operating characteristics; SVC_JM, joint model established by support vector machines classifier; and SVC_RM, radiomic model established by support vector machines classifier.

Figure 4. Importance of predictor variables based on the SHAP algorithm for the LRC model.

A, Importance ranking plot of variables for LRC_JM model in diagnosing PH. B, Beeswarm plots of the LRC_JM model in diagnosing PH. C, Importance ranking plot of variables for LRC_JM model in diagnosing PAH. D, Beeswarm plots of the LRC_JM model in diagnosing PAH. E, Individual explanation for an NPH sample. F, Expression heat map of radiomic variables for LRC_JM model of PH and NPH samples. G, Individual explanation for a PH sample. Elg indicates elongation; GNU, GrayLevel NonUniformity; HHH_SRHGE, Short Run High GrayLevel Emphasis after High (x) ‐ High (y) ‐ High (z) pass wavelet filter; (F), HLL_SAE, Small Area Emphasis after High (x) ‐ Low (y) ‐ Low (z) pass wavelet filter; HHL_SAHGE, Small Area High GrayLevel Emphasis after High (x) ‐ High (y) ‐ Low (z) pass wavelet filter; LA, left atrial diameter measured by echocardiology; LHH_Kur, kurtosis after Low (x) ‐ High (y) ‐ High (z) pass wavelet filter; LHL_DA, difference average after Low (x) ‐ High (y) ‐ Low (z) pass wavelet filter; LHL_Mean, mean after Low (x) ‐ High (y) ‐ Low (z) pass wavelet filter; LHL_Min, minimum after Low (x) ‐ High (y) ‐ Low (z) pass wavelet filter; LLL_ BSY, busyness after Low (x) ‐ Low (y) ‐ Low (z) pass wavelet filter; LLL_Coa, coarseness after Low (x) ‐ Low (y) ‐ Low (z) pass wavelet filter; LLL_Cor, correlation after Low (x) ‐ Low (y) ‐ Low (z) pass wavelet filter; LLL_SZNUN, Size Zone NonUniformity Normalized after Low (x) ‐ Low (y) ‐ Low (z) pass wavelet filter; Log_Imcl, Imcl after Laplacian of Gaussian filter; LRC_JM, joint model established by logistic regression classifier; NPH, nonpulmonary hypertension; PA/AO, ratio of pulmonary artery width and aortic width measured by computed tomography; PA_E, inner diameter of pulmonary artery trunk measured by echocardiology; PH, pulmonary hypertension; RV, transverse diameter of right ventricle; SHAP, Shapley additive explanations; and TRV, tricuspid regurgitation velocity.

Figure 5. Model performance of radiomic models and joint models in classifying PAH.

A, ROC curve of 6 models in the resampled balanced developed data. B, Calibration curve of 6 constructed models. C, PR curve of 6 models in a raw developed cohort. D, PR curve of 6 constructed models in a validated cohort. LRC_JM indicates joint model established by logistic regression classifier; LRC_PA/AO, computed tomography‐measured diameter ratio of the main pulmonary artery and the aorta diagnostic model established by logistic regression classifier; LRC_RM, radiomic model established by logistic regression classifier; LRC_TRV, echocardiographic measured tricuspid regurgitation velocity diagnostic model established by logistic regression classifier; PAH, pulmonary arterial hypertension; PR, precision‐recall; RFC_JM, joint model established by random forests classifier; RFC_RM, radiomic model established by random forests classifier; ROC, receiver operating characteristics; SVC_JM, joint model established by support vector machines classifier; and SVC_RM, radiomic model established by support vector machines classifier

Figure 6. Predictor variables and Kaplan–Meier analysis of PH prognosis.

A, Radiomic features were selected based LASSO and multivariate Cox for PH prognosis. B, Kaplan–Meier analysis of RM and JM models in the risk of adverse events in derivation cohort. C, Radiomic and clinical features were selected based LASSO and multivariate Cox for PH prognosis. D, Kaplan–Meier analysis of RM and JM models in the risk of adverse events in validation cohort. AUC indicates area under the curve; HLH, High (x) ‐ Low (y) ‐ High (z) pass wavelet filter; LLH, Low (x) ‐ Low (y) ‐ High (z) pass wavelet filter; LLL, Low (x) ‐ Low (y) ‐ Low (z) pass wavelet filter; MCC, Maximal Correlation Coefficient; JM, joint models; LASSO, least absolute shrinkage and selection operator; LLH, LLL, MCC, PA, diameter of pulmonary artery trunk measured by computed tomography; PH, pulmonary hypertension; RM, radiomic models; and RV, transverse diameter of right ventricle measured by echocardiology.

Figure 7. Model performance of radiomic models and joint models in PH prognosis.

A, ROC curve of RM, JM models and the standard fourth‐layer risk stratification in derivation cohort. B, Time‐AUC of RM, JM, and the standard fourth‐layer risk stratification in derivation cohort. C, ROC curve of RM and JM models and the standard fourth‐layer risk stratification in validation cohort. D, Time‐AUC of RM, JM, and the standard fourth‐layer risk stratification in validation cohort. The CIs for the time‐dependent AUC in derivation cohort were estimated using the 10×5‐fold cross‐validation approach while in validation cohort using asymptotic normality assumption. AUC indicates area under the curve; JM, joint models; PH, pulmonary hypertension; RM, radiomic models; ROC, receiver operating characteristic; and SCORE_ESC, the score calculated by simplified 4‐strata risk assessment tool of the 2022.