Clinical application of artificial intelligence-assisted pulmonary vascular imaging in pulmonary hypertension associated with left heart disease

Background Pulmonary hypertension associated with left heart disease (PH-LHD) is the most prevalent PH subtype1. While right heart catheterization (RHC) remains the gold standard, its invasiveness and inability to detect early microvascular changes limit clinical utility.2 Although CTA reveals parameters-vascular remodeling correlations,3 quantification of hemodynamic abnormalities and microvascular lesions remains underexplored. Current RHC-based assessments fail to characterize progression due to impractical tissue sampling, hindering early clinical intervention in PH-LHD.4 This study aimed to develop an AI-driven radiomics platform for automated pulmonary vascular reconstruction and to evaluate whether imaging signatures could identify early microvascular remodeling preceding hemodynamic changes,5 further establish a multi-scale predictive model integrating imaging signatures with pathological progression. Methods We developed an AI-radiomics platform with a 3D Slicer plugin for automated pulmonary vascular analysis. A prospective study (healthy/PAH/PH-LHD cohorts) underwent CTA, echocardiography, and RHC. The platform employed deep learning and optimized GrowCut algorithms to reconstruct eight-branch vascular networks with automated artery/vein classification, using Vascular Modeling Toolkit (VMTK)-based centerline extraction. Radiomics features underwent Boruta selection to identify remodeling biomarkers, validated through machine learning models.7 Preclinical validation used monocrotaline (MCT)-induced PAH and transverse aortic constriction (TAC)-induced PH-LHD rat models with hemodynamic/histopathological assessments. Cross-species analysis correlated radiomics-predicted remodeling scores with histomorphometric data. Results The AI-radiomics platform enabled automated pulmonary vascular analysis through high-precision 3D vessel segmentation, reconstructing eight-order branching networks at submillimeter resolution. Clinically, radiomics-derived venous tortuosity and fractal dimension differentiated PH-LHD from PAH, outperforming RHC metrics, while venous dominance identified Group 2 PH pre-hemodynamic confirmation. Preclinical validation in PH-LHD rats revealed micro-CT-detected early vascular remodeling at 6 weeks (without RVSP elevation), showing increased vessel density and venous tortuosity. Branching complexity and tortuosity correlated with histomorphometric parameters, validating their utility in assessing remodeling. Conclusion This study developed a 3D Slicer plugin with automated template deformation and deep learning segmentation for AI-assisted pulmonary vascular reconstruction and radiomics extraction. We identified pulmonary venous tortuosity and vascular branch grading as sensitive predictors of early vascular remodeling in PH-LHD, offering valuable and portable methods for early clinical intervention and management.