Identification of pulmonary hypertension with pulmonary vascular surface volume ratio by non-contrast CT Authors: Qing Yu, Yating Zhao, Yini Huang, Mingya Zhang, Gulinigaer Wufulihasimu, Yijia Zhou, JiaJia Zhang, Xiaopeng Xu, Wei Zhang, Yanxi Zeng, Wenhui Peng, Yunshan Cao Published: 2026 ABSTRACT AbstractBackgroundThe diagnosis of pulmonary hypertension (PH) relies primarily on invasive right heart catheterization (RHC). Although imaging modalities, including echocardiography, computed tomography angiography (CTA) and computed tomography (CT), offer clinical utility, their effectiveness in early detection remains limited.ObjectivesTo clarify the relationship between group 1/2 PH and radiomic-morphological features of the pulmonary vasculature derived from non-contrast CT.MethodsWe retrospectively enrolled control patients and patients with group 1 PH (PAH) and group 2 PH (PH-LHD) from two centers. All participants underwent RHC and chest CT. We extracted the radiomic features from pulmonary arteries and veins, identifying key radiomics markers and further exploring their association with early pathological changes through animal experiments. Subsequently, we analyzed morphological characteristics from 95,197 vascular centerlines. Multivariable models were employed to evaluate their efficacy in identifying PH and discriminating PH-LHD in validation cohort.ResultsThe training cohort included 49 patients (20 controls, 13 with PAH, 16 with PH-LHD), and validation cohort included 30 patients. Among 212 radiomics features analysis, the pulmonary arterial surface volume ratio (SVR) correlated strongly with mean pulmonary artery pressure (mPAP) (r = -0.602, P < 0.001), while the venous SVR showed a strong correlation with pulmonary artery wedge pressure (PAWP) (r = -0.651, P < 0.001). In animal models, venous SVR demonstrated the ability to reflect early pulmonary venous remodeling in PH-LHD prior to hemodynamic changes. Morphologically, PH-LHD patients showed increased distal vessel tortuosity with reduced torsion and curvature (all P < 0.001). SVR and models combining morphological features demonstrated promising performance for identifying PH (AUC 0.895 (95%CI (0.731-1.000)) vs. 0.890 (95%CI (0.736-1.000)) and distinguishing PH-LHD (AUC 0.868 (95%CI (0.672-1.000) vs. 0.846 (95%CI 0.639–0.990)) in validation cohort.ConclusionCT-based radiomic-morphological analysis provides a basis for identifying PH and distinguishing its subtypes. Specifically, arterial SVR aids in PH detection, while venous SVR helps differentiate PH-LHD from PAH, together forming a clinically useful screening pathway. Preclinical evidence additionally links venous SVR to venous remodeling that may precede hemodynamic changes. These findings require validation in larger PH populations.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s12872-026-06078-z.Keywords: Pulmonary hypertension, Surface volume ratio, Radiomics, CT imaging Introduction Pulmonary hypertension (PH) is a major global health issue affecting all age groups, with a prevalence of approximately 1% worldwide. Its insidious onset and late clinical presentation hinder early detection [1]. The current definitive diagnosis relies on right heart catheterization (RHC), which is an invasive procedure that limits its widespread use [2]. Although echocardiography remains the most common screening tool for PH, its operator dependence and inability to differentiate subgroups limit its utility [1]. Similarly, although computed tomography angiography (CTA) enables an objective assessment of the pulmonary vasculature and improves the identification of PH, its dependence on contrast agents limits its use in population screening [3]. In contrast to both, non-contrast CT provides an objective and practical method for lung disease screening while also enabling vascular assessment. Therefore, we aim to investigate the relationship between CT-derived features and PH. A large number of studies have demonstrated the potential of CT for non-invasive PH assessment, with advances in imaging analysis enabling the detection of early vascular alterations [4]. Radiomic parameters reveal topological features reflecting microstructural changes [5], and centerline analysis quantifies subtle vascular distortions associated with histological wall remodeling [6]. These methods can extract numerous CT features to effectively assess patient prognosis across various diseases, including PH [7, 8]. However, current CT-based assessment of PH has focused on pulmonary artery analysis and estimation of mean pulmonary arterial pressure (mPAP), while overlooking venous characteristics and the key indicator of typing: pulmonary arterial wedge pressure (PAWP) [9, 10]. This gap is critical since PH subgroups exhibit distinct pathophysiological and hemodynamic characteristics, with group 1 and group 2 being the most representative. Group 1 (pulmonary arterial hypertension (PAH)), the most classic form of PH, is characterized by constriction and remodeling of small pulmonary arteries, resulting in elevated vascular resistance [1]. In contrast, group 2 (pulmonary hypertension associated with left heart disease (PH-LHD)) is the most prevalent form, primarily driven by elevated left atrial pressure causing pulmonary venous congestion [11]. These differences underlie their divergent pathological profiles and treatment strategies [12]. In particular, pulmonary venous remodeling represents a characteristic feature of PH-LHD and may precede arterial changes. This pattern distinguishes PH-LHD from the arterial remodeling typical of PAH, suggesting that venous characteristics may support subtype classification and help guide therapeutic decisions [13]. Accordingly, we will explore whether there are differences in pulmonary vascular characteristics between PAH and PH-LHD, and attempt to identify key parameters capable of detecting early vascular remodeling and distinguishing PAH and PH-LHD. Methods Patient inclusion and exclusion This retrospective study was approved by the Ethics Committee of Gansu Provincial Hospital (Approval No. 2024 − 521) and the Ethics Committee of Sichuan Provincial People’s Hospital (Approval No. 2024 − 469). The informed consent was waived in both centers. In accordance with Helsinki Declaration principles, consecutive PH patients were enrolled from January 2020 to January 2024. Inclusion criteria were mPAP > 20 mmHg and PVR > 2 WU. PAH (group 1 PH) was defined as having a PAWP of ≤ 15 mmHg, while PH-LHD (group 2 PH) was defined as having a PAWP of > 15 mmHg. Risk stratification was based on the 2022 ESC/ERS guidelines, combining WHO-FC, 6-minute walk distance (6MWD), and NT-proBNP levels [1]. For analysis, intermediate-low and intermediate-high risk were merged into an “Intermediate-low/high” group. All enrolled patients underwent CT and RHC within one week of each other. For PAH and PH-LHD patients, only treatment‑naïve patients were enrolled. The control group comprised patients who were initially suspected of having PH but were ultimately confirmed not to have PH following a definitive RHC assessment. Their primary comorbidities included chronic cardiopulmonary diseases, such as hypertension (30%), stable coronary artery disease (10%), arrhythmia (10%), and chronic obstructive pulmonary disease (COPD) or relevant lung disease (5%). The enrolled PAH patients included the following subtypes: idiopathic PAH (n = 4, 30.8%), connective tissue disease-associated PAH (n = 8, 61.5%), and portopulmonary hypertension (n = 1, 7.7%). Exclusion criteria included severe hepatic or renal impairment, significant valvular stenosis and group 3, 4 or 5 PH. Group 3 PH (PH associated with lung diseases and/or hypoxia) was excluded based on the presence of overt lung disease on CT, whereas occult lung disease was not excluded. Group 4 PH (PH associated with pulmonary artery obstructions) was excluded based on CTA evaluation. Comprehensive baseline data were collected. In addition, we retrospectively collected an external validation set, applying the same inclusion and exclusion criteria as described previously. Chest CT radiomics-centerline feature extraction Non-contrast chest CT scans were performed from the apex to the diaphragm using United Imaging scanners (120 kVp, 300 mA, 0.3 s rotation, 64 × 0.625-mm collimation). By referencing both the foundational nnU‑Net framework and related methodological advances in pulmonary vessel segmentation [14, 15], we constructed a dataset of 1000 annotated cases reviewed by two radiologists. A five-fold iterative training strategy was employed, where the model was progressively refined by predicting unannotated cases, correcting the segmentations, and retraining until all cases were included. To enhance cross-center generalizability, we subsequently performed transfer learning using 100 additional cases from a second center with diverse CT scanners and acquisition parameters. The final model was ensembled from five folds and validated on 40 unseen cases equally drawn from both centers, demonstrating robust cross-center segmentation performance. The model achieved high performance (arteries: DSC 0.912 ± 0.054; veins: DSC 0.936 ± 0.016), with multi-center training improving DSC by 0.5–1.9%, and demonstrating robust cross-center applicability (Table 1). Radiomic features were extracted using PyRadiomics, including morphological, topological, textural, and functional parameters [16]. Vascular centerlines and geometric properties (radius, torsion, curvature, tortuosity and length) were derived using the Vascular Modeling Toolkit (VMTK) [17]. The branching hierarchy was defined from level 0 (main trunk), with vessels categorized as proximal (level 0–4), medium (level 5–9), or distal (level ≥ 10) based on their branching level. Table 1Segmentation performance of the optimized nnU-Net modelScenarioPulmonary Arteries (DSC/IOU)Pulmonary Veins(DSC/ IOU)Airways(DSC/ IOU)Single-Center 5- Fold0.9069 ± 0.0689 /0.8357 ± 0.10610.9177 ± 0.0508 /0.8514 ± 0.08160.9265 ± 0.0490 /0.8662 ± 0.0782Multi-Center Transfer0.9118 ± 0.0539 /0.8416 ± 0.08500.9363 ± 0.0163 /0.8807 ± 0.02860.9275 ± 0.0367 /0.8667 ± 0.0602Performance is evaluated using Dice Similarity Coefficient (DSC) and Intersection over Union (IOU) across single-center 5-fold cross-validation and independent multi-center transfer learning scenarios Segmentation performance of the optimized nnU-Net model 0.9069 ± 0.0689 / 0.8357 ± 0.1061 0.9177 ± 0.0508 / 0.8514 ± 0.0816 0.9265 ± 0.0490 / 0.8662 ± 0.0782 0.9118 ± 0.0539 / 0.8416 ± 0.0850 0.9363 ± 0.0163 / 0.8807 ± 0.0286 0.9275 ± 0.0367 / 0.8667 ± 0.0602 Performance is evaluated using Dice Similarity Coefficient (DSC) and Intersection over Union (IOU) across single-center 5-fold cross-validation and independent multi-center transfer learning scenarios Right heart catheterization All right heart catheterization (RHC) procedures were conducted by an interventional cardiologist with over five years’ experience in this specialty, in accordance with established clinical protocols [18]. Standard hemodynamic parameters were systematically obtained, including superior/inferior vena cava pressure, right atrial pressure (RAP), right ventricular pressure (RVP), systolic/diastolic pulmonary artery pressure (sPAP, dPAP), mean pulmonary artery pressure (mPAP), and pulmonary arterial wedge pressure (PAWP). PAWP was measured using a balloon-tipped catheter, and the reported value represents the average of three end-expiratory measurements taken at the mid-A wave. Mixed venous oxygen saturation (SvO₂) was measured, and cardiac output (CO) was determined via the thermodilution method. Cardiac index (CI) was derived by indexing CO to body surface area (BSA). BSA was calculated using the Du Bois formula. Pulmonary vascular resistance (PVR) in Wood units: PVR = (mPAP – PAWP) / CO. The operators performing RHC were blinded to all echocardiographic findings. Animals and PH models All animal procedures followed the Guide for the Care and Use of Laboratory Animals (National Institutes of Health, Bethesda, MD, USA) and were approved by the Institutional Animal Care and Use Committee (IACUC) of Tongji University (Shanghai, China) (Approval No. SHDSYY-2024-4414). The study complied with American Physiological Society guidelines and used only male Sprague-Dawley (SD) rats (GemPharmatech Co., Ltd., Nanjing, China) to minimize the effects of sex hormones. Monocrotaline (MCT)-induced PAH rats received one intraperitoneal injection (60 mg/kg, Sigma-Aldrich) [19]. PH-LHD rats were established by transverse aortic constriction (TAC). Anesthetized rats (pentobarbital, 50 mg/kg, i.p.; Sigma-Aldrich) underwent thoracotomy, where the thoracic aorta was ligated with a 7 − 0 silk suture against a 22-gauge needle [20]. Sham controls had identical surgery without ligation. The success of the MCT-induced PAH and TAC-induced PH-LHD models was confirmed by echocardiography and RHC at 4 weeks and 8 weeks post-modeling. Echocardiography and hemodynamic measurements Transthoracic echocardiography was performed under isoflurane anesthesia (induction: 5% maintenance; 2% RWD system) using a Vevo 2100 platform (VisualSonics) equipped with a 250 MHz transducer. Standard views were acquired, including the parasternal long-axis, the left ventricular short-axis transitioning to the pulmonary artery, the right parasternal long-axis and the apical four-chamber view [21]. The measured parameters included left ventricular ejection fraction (LVEF), fractional shortening (FS%), pulmonary artery acceleration time (PAT), ejection time (PET), estimated pulmonary artery systolic pressure (PASP), end-diastolic right ventricular (RV) inner diameter, RV wall thickness, tricuspid annular plane systolic excursion (TAPSE), and calculated RV ejection fraction (RVEF). Data analysis utilized EchoPAC software. Micro-CT scanning and right heart catheterization Under pentobarbital sodium anesthesia (50 mg/kg, i.p.), supine rats underwent jugular vein catheterization using a heparinized 24G catheter (100 IU/mL). The catheter delivered iodixanol contrast (1:1 saline, total 1 mL; Hengrui Medicine) at a rate of 25 µL/min during micro-CT scanning (80 kV, 200 µA, 112 s acquisition, 56 ms exposure, 60 μm resolution, 2000 projections). The images were reconstructed with InViscan Recon Software (V3.6) and 3D Slicer. Post-scan right heart catheterization under 2% isoflurane anesthesia measured right ventricular systolic pressure (RVSP) via a Millar catheter connected to an ADV500 system [22]. Histology and immunohistochemical staining The lung specimens underwent fixation in 10% neutral buffered formalin, followed by dehydration, paraffin embedding, and sectioning at 5 μm thickness. The sections were stained with Hematoxylin and Eosin (H&E), EVG and Picrosirius Red Staining for general morphology and elastic fiber visualization. EVG staining provided additional discrimination based on the characteristic morphology and pattern of elastic fibers. Vascular remodeling parameters were quantified as follows: Medial wall thickness percentage WT% = [(External Diameter - Internal Diameter) / External Diameter] × 100 and vessel wall area percentage WA% = [(External Area - Internal Area) / External Area] × 100. The extent of lung fibrosis was also determined by calculating the percentage area of fibrotic lesions within examined fields. Diameters of all discernible pulmonary arteries and veins within microscopic fields were recorded. Statistical analysis Data are presented as mean ± SD or median (p25, p75) for continuous variables and absolute numbers (percentages) for categorical variables. Group differences for continuous variables were assessed by one-way ANOVA followed by Tukey’s post hoc test for normally distributed data, or by Kruskal-Wallis H test followed by Dunn’s multiple comparison test for non-normally distributed data. For categorical variables, group differences were assessed by the chi-square test, followed by pairwise chi-square tests with Bonferroni correction. Correlations were assessed using Spearman’s analysis. The initial screening of radiomic features was exploratory in nature, and therefore no correction for multiple comparisons was applied at this stage. Statistical significance was defined as P < 0.05. Analyses were performed using GraphPad Prism v9, IBM SPSS Statistics, and Python 3.9. Results Patients of PH subgroups enrollment and baseline characteristics We conducted a retrospective clinical study. A total of 49 patients were enrolled as a training set based on stringent inclusion and exclusion criteria, comprising control subjects (n = 20), patients with PAH (n = 13) and PH-LHD (n = 16). In addition, we enrolled another 30 patients from the second center as a validation set, comprising control subjects (n = 10), patients with PAH (n = 13) and PH-LHD (n = 7). Comprehensive baseline data of the training set were collected (Table S1). Hemodynamically and in terms of cardiac function, both the PAH and PH-LHD groups exhibited elevated mPAP and PVR, accompanied by right heart dilation and functional impairment. However, the PH-LHD group was additionally characterized by a significantly enlarged left atrium and elevated PAWP, which distinguishes it from other PH subgroups. It is worth noting that no significant differences in LVEF were observed among the three groups, suggesting that the PH-LHD patients predominantly had heart failure with preserved ejection fraction (HFpEF) (Table 2). Table 2Baseline characteristics of the enrolled cohortParameterControlPAHP Value(vs. Control)PH-LHDP Value(vs.Control)P Value(vs. PAH)Age53.9 ± 11.447.2 ± 13.40.11755.2 ± 15.30.5450.091Gender (Male)10 (50.0%)3 (23.1%)7 (43.8%)(Female)10 (50.0%)10 (76.9%)0.2379 (56.2%)0.9700.433RHCmPAP, mmHg15.6 ± 2.645.8 ± 17.6< 0.001*42.1 ± 10.3< 0.001*0.517PVR, WU1.2 (1.1, 1.6)6.4 (3.9, 13.9)< 0.001*5.0 (2.2, 6.1)< 0.001*0.041*PAWP, mmHg8.0 (8.0, 10.5)7.0 (6.0, 9.0)0.14920.0 (18.5, 21.8)< 0.001*< 0.001*Risk stratification & factorsRisk-assessment Low20 (100.0%)5 (38.5%)0 (0.0%) Intermediate-low/high0 (0.0%)5 (38.5%)9 (56.2%) High0 (0.0%)3 (23.1%)< 0.001*7 (43.8%)< 0.001*0.023*WHO-FC I19 (95.0%)1 (7.7%)0 (0.0%) II1 (5.0%)8 (61.5%)4 (25.0%) III0 (0.0%)3 (23.1%)10 (62.5%) IV0 (0.0%)1 (7.7%)< 0.001*2 (12.5%)< 0.001*0.1036MWD, m483.4 ± 40.1360.2 ± 56.9< 0.001*336.8 ± 99.5< 0.001*0.435proBNP, pg/mL53 (32, 58)313 (84, 738)< 0.001*697 (451, 2120)< 0.001*0.069EchocardiographyRA enlarged No15 (75.0%)3 (23.1%)5 (31.2%) Yes5 (25.0%)10 (76.9%)0.010*11 (68.8%)0.022*0.697RV enlarged No16 (80.0%)3 (23.1%)5 (31.2%) Yes4 (20.0%)10 (76.9%)0.004*11 (68.8%)0.009*0.697LA enlarged No14 (70.0%)8 (61.5%)5 (31.2%) Yes6 (30.0%)5 (38.5%)0.71311 (68.8%)0.048*0.209LV enlarged No17 (85.0%)11 (84.6%)10 (62.5%) Yes3 (15.0%)2 (15.4%)1.0006 (37.5%)0.14610.238MPAD, mm23.75 ± 2.1734.00 ± 6.43< 0.001*30.25 ± 4.07< 0.001*0.083TAPSE, mm23 (21, 25)20 (15, 21)0.036*16 (14, 19)0.001*0.201PASP, mmHg16 (15, 17)74 (55, 86)< 0.001*56 (45, 76)< 0.001*0.210LVEF, %65.0 (61.8, 65.0)65.0 (62.0, 70.0)0.22263.5 (46.0, 66.5)0.4500.124Data are presented as mean ± SD or median (p25, p75) for continuous variables and absolute numbers (percentages) for categorical variables. Group differences for continuous variables were assessed by one-way ANOVA followed by Tukey’s post hoc test for normally distributed data, or by Kruskal-Wallis H test followed by Dunn’s post hoc test with Bonferroni correction for non-normally distributed data. For categorical variables, group differences were assessed by the chi-square test, followed by pairwise chi-square tests with Bonferroni correction for multiple comparisons RHC right heart catheterization, mPAP mean pulmonary arterial pressure, PVR pulmonary vascular resistance, PAWP pulmonary arterial wedge pressure, WHO-FC world health organization functional classification, 6MWD 6-minute walk distance, proBNP pro-B-type natriuretic peptide, MPAD main pulmonary artery diameter, RA/RV/LA/LV enlarged right atrial/ventricular or left atrial/ventricular enlargement, TAPSE tricuspid annular plane systolic excursion, PASP pulmonary arterial systolic pressure, LVEF left ventricular ejection fractionA two-sided P < 0.05 was considered statistically significant Baseline characteristics of the enrolled cohort Data are presented as mean ± SD or median (p25, p75) for continuous variables and absolute numbers (percentages) for categorical variables. Group differences for continuous variables were assessed by one-way ANOVA followed by Tukey’s post hoc test for normally distributed data, or by Kruskal-Wallis H test followed by Dunn’s post hoc test with Bonferroni correction for non-normally distributed data. For categorical variables, group differences were assessed by the chi-square test, followed by pairwise chi-square tests with Bonferroni correction for multiple comparisons RHC right heart catheterization, mPAP mean pulmonary arterial pressure, PVR pulmonary vascular resistance, PAWP pulmonary arterial wedge pressure, WHO-FC world health organization functional classification, 6MWD 6-minute walk distance, proBNP pro-B-type natriuretic peptide, MPAD main pulmonary artery diameter, RA/RV/LA/LV enlarged right atrial/ventricular or left atrial/ventricular enlargement, TAPSE tricuspid annular plane systolic excursion, PASP pulmonary arterial systolic pressure, LVEF left ventricular ejection fraction A two-sided P < 0.05 was considered statistically significant Surface volume ratio serves as a potential predictor among radiomics features A self-developed cross-center nnU-Net model was used to automate vasculature segmentation, artery-vein separation, and feature extraction from non-contrast CT scans. The most relevant radiomic biomarkers were identified by extracting features that quantified intensity, texture, and spatial patterns. The surface volume ratio (SVR) exhibited the strongest overall correlations. Arterial SVR correlated with mPAP (r = -0.602, P < 0.001) and risk assessment (r = -0.684, P < 0.001), while venous SVR was closely linked to PAWP (r = -0.651, P < 0.001), the key diagnostic criterion for PH-LHD (Table 3). SVR was calculated as Surface Area / Mesh Volume over the entire arterial or venous mask. Physiologically, a decreased SVR corresponds to increased vascular volume, indicating the extent of overall vascular dilation and pressure levels [23]. Consistent with this, 3D vascular models revealed arterial dilation accompanied by decreased SVR in both PAH and PH-LHD groups compared to the control group. Notably, the PH-LHD group exhibited more pronounced venous dilation and a greater reduction in SVR than both control and PAH groups (Fig. 1). While previous studies emphasized its diagnostic value [24], our findings suggested that the surface-to-volume ratio may offer superior performance in screening PH. Table 3Top 5 radiomic features correlated with hemodynamics and risk assessmentParameterr valueP_valuemPAP Arterial Surface Volume Ratio-0.602< 0.001* Arterial Mesh Volume0.572< 0.001* Arterial Voxel Volume0.572< 0.001* Arterial First Order 10 Percentile0.512< 0.001* Arterial Large Area High Gray Level Emphasis0.481< 0.001*PAWP Venous Surface Volume Ratio-0.651< 0.001* Venous GLRLM Gray Level Non-Uniformity0.600< 0.001* Venous Size Zone Non-Uniformity Normalized-0.567< 0.001* Venous Small Area Emphasis-0.566< 0.001* Venous GLDM Gray Level Non-Uniformity0.566< 0.001*Risks assessment Arterial Surface Volume Ratio-0.684< 0.001* Venous Difference Variance-0.554< 0.001* Arterial Mesh Volume0.545< 0.001* Venous Surface Volume Ratio-0.542< 0.001* Arterial Voxel Volume0.540< 0.001*mPAP mean pulmonary arterial pressure, PAWP pulmonary arterial wedge pressure*P < 0.05: The correlation assessed by Spearman’s correlation analysis was significant Top 5 radiomic features correlated with hemodynamics and risk assessment mPAP mean pulmonary arterial pressure, PAWP pulmonary arterial wedge pressure *P < 0.05: The correlation assessed by Spearman’s correlation analysis was significant Radiomic characteristics of pulmonary arteries and veins in human subjects. A Pulmonary artery and pulmonary vein reconstruction images and vascular changes (arrow) of control, PAH, PH-LHD group. B Comparison of the arterial and venous SVR. Statistical significance was determined by One-way ANOVA or Kruskal-Wallis H test with Tukey’s post hoc analysis SVR reflected early pulmonary vascular remodeling in rat models To elucidate the relationship between pulmonary vascular SVR and pathological remodeling in PH, we established the MCT-induced PAH and TAC-induced PH-LHD models, comparing their micro-CT and pathological characteristics (Fig. 1A). Comprehensive hemodynamic analysis confirmed the presence of PH in the MCT-induced and late-stage (8-week) TAC models, characterized by elevated RVSP and echocardiographic alterations (Fig. 2B–C, Table S2). In contrast, early-stage (4-week) TAC animals maintained normal hemodynamic parameters, yet demonstrated significant venous remodeling, characterized by increased wall thickness, cross-sectional area and fibrosis, particularly evident in small-to-medium vessels (< 100 μm) (Fig. 2B-F, Figures S1, Table S2). To evaluate whether pulmonary vascular SVR could specifically reflect pathological remodeling, we compared SVR across groups and observed that changes in SVR consistently aligned with pathological alterations (Fig. 2G). Notably, venous SVR was closely associated with both venous remodeling and arterial fibrosis. It effectively distinguished early-stage PH-LHD from controls and PAH (P = 0.038), whereas arterial SVR did not (P = 0.616) (Fig. 2H). These results confirmed that in rat models venous SVR serves as a sensitive marker of early venous remodeling and a promising feature for early PH-LHD identification. Micro-CT and histopathological validation of pulmonary arteries and veins in rat models of PH. A Schematic of the experimental design, from model establishment to histological analysis. B-C Hemodynamic and echocardiographic parameters including RVSP, LVEF and TAPSE. D Representative hematoxylin and eosin (H&E) staining of pulmonary arteries and veins across different diameter ranges (< 50 μm, 50–100 μm, ≥ 100 μm) from all experimental groups. Scale bar, 50 μm. (E-F) Quantitative analysis of medial wall thickness (WT%) for pulmonary arteries (E) and veins (F). G Comparison of SVR in control and PH groups. Correlation was assessed using Spearman’s rank correlation coefficient. H Correlation matrix between histopathological indices and SVR. Statistical significance was determined by one-way ANOVA or Kruskal-Wallis H test with Tukey’s post hoc analysis. WT = wall thickness, WA = wall area, FA = fibrotic area Centerline parameters reveal morphological differences between PH subgroups As additional morphological validation [25], we further extracted 95,197 centerlines from patient vascular models and categorized them as proximal large vessels (level 0–4), medium vessels (level 5–9), and distal vessels (level ≥ 10) (Figs. 3 and 4). Compared with the control group, both PH subgroups exhibited shared arterial abnormalities, including significant main pulmonary artery dilation, an increased proportion of distal arteries, and a higher percentage of highly tortuous vessels (tortuosity > 0.1) (Table S3). In contrast, the PH-LHD group demonstrated a distinct and specific venous pattern characterized by a reduced proportion of distal veins and a loss of small vessels (diameter < 1 mm), which significantly differentiated it from both the control group and the PAH group (all P < 0.001) (Table S4). We quantified vascular morphology across radius and branching levels, correlating features with hemodynamics and risk. The top 5 correlated morphological features linked to mPAP were all related to arterial tortuosity (Table S5), suggesting that abnormal pulmonary arterial pressure primarily alters vascular curvature and validating a prior observation [26]. Workflow of the nnU-Net-based AI platform extraction of the vascular centerline network Flowchart of centerlines quality control CT‑based screening and subtyping of PAH/PH-LHD using SVR and morphological features Based on correlation and LASSO analysis (Figure S2), selected radiomic and morphological features were used to develop predictive models. To validate the predictive performance of SVR and its regression model, we assessed an external validation cohort (n = 30) (Table S6). While initial stepwise multivariable linear regression models for continuous mPAP and PAWP prediction showed limited performance (Table S7). We therefore developed logistic regression models for binary classification tasks including Control vs. PH, PAH vs. PH-LHD, and risk stratification. These models all incorporated SVR and demonstrated robust diagnostic accuracy (Fig. 5). Notably, the mPAP model covariates were exclusively arterial, whereas the PAWP model covariates were solely venous, which highlights subgroup-specific pathophysiological changes. In the training cohort, CT-based diagnosis showed comparable performance to the echocardiographic TAPSE/PASP ratio for PH detection (AUC 0.948 (95% CI 0.887–0.993) vs. 0.947 (95% CI 0.884–0.990)) and risk stratification (AUC 0.940 (95% CI 0.868–0.990) vs. 0.870 (95% CI 0.744–0.957)). Importantly, the CT model demonstrated significantly superior efficacy in differentiating PAH from PH-LHD (AUC 0.870 (95% CI 0.721–0.972) vs. 0.538 (95% CI 0.360–0.720), P = 0.011) (Fig. 5 and Table S8). This highlights the importance of pulmonary venous features in subtype classification. The combined regression model demonstrated superior performance to SVR alone in both the identification of PH (AUC: 0.948 vs. 0.890) and the discrimination of PH-LHD (AUC: 0.870 vs. 0.841) (Fig. 5). Multivariable Logistic Regression Model in Training set. ROC curves and corresponding AUC for surface volume ratio (SVR) and its predictive logistic model incorporating other features. A Identifying PH: P = 1 / (1 + e⁻ᶻ), z = 1.120–1.221 × Arterial SVR + 0.997 × Arterial Mesh Volume + 0.698 × Arterial Mean Tortuosity (Medium Vessel) + 0.910 × Arterial Mean Tortuosity (Range: 0.05–0.1). B Discriminating PAH from PH-LHD: P = 1 / (1 + e⁻ᶻ), where z = 0.183–0.833 × Venous SVR + 0.570 × Venous Gray Level Non-Uniformity – 0.351 × Venous Low Gray Level Run Emphasis – 0.502 × Venous Size Zone Non-Uniformity Normalized. C Risk stratification: P = 1 / (1 + e⁻ᶻ), z = 0.001–1.456 × Arterial SVR + 0.988 × Arterial Mean Radius (Proximal Vessel) – 0.525 × Venous Difference Variance + 0.333 × Arterial Mesh Volume In the validation cohort, the results indicated that both SVR and the regression model continued to demonstrate high diagnostic efficacy (Fig. 6; Table 4). Specifically, SVR alone maintained robust diagnostic performance, with all AUCs above 0.800 (Fig. 6A-C; Table 4) and accuracy above 0.750 (Fig. 6D-E; Table 4). In a supplementary analysis comparing PH-LHD to PAH and controls combined, the CT model showed significantly superior discrimination to TAPSE/PASP in the validation cohort (AUC 0.876 (95% CI: 0.714–0.995) vs. 0.609 (95% CI 0.348–0.836), P = 0.023) (Figure S3A-B). A sensitivity analysis excluding patients with COPD and relevant lung disease from the control group did not materially alter these findings (Figure S3C-D). Additionally, although current guidelines recommend left atrial (LA) volume for diastolic dysfunction assessment, the lack of LA anterior-posterior diameter (LAd) precluded its formal calculation. Our estimated LA volume (LAVe) still showed inferior discriminative ability between PAH and PH-LHD compared to pulmonary venous SVR (Figure S4) [27]. The optimal diagnostic thresholds were confirmed at arterial SVR < 0.687 for identifying PH and venous SVR < 0.669 for discriminating PH-LHD (Table 4). The diagnostic efficacy of SVR and the Regression Model in Validation set. A-C The ROC curves of arterial/venous SVR and the two regression models in the validation set for the diagnosis of PH (A), PH-LHD (B) and intermediate/high risks of PH (C). D the confusion matrix of the regression model in the training set for the diagnosis of PH, PH-LHD, and intermediate/high risks of PH in the validation set, with the accuracy of 0.867, 0.800, and 0.800, respectively. E the confusion matrix of arterial and venous SVR for the diagnosis of PH, PH-LHD, and intermediate/high risks of PH in the validation set, with the accuracy of 0.900, 0.750, and 0.867, respectively Table 4The predictive performance of SVR and regression models in validation setControl vs. PH patientsPAH vs. PH-LHD patientsLow vs. Intermediate/High risksRegression ModelArterial SVRRegression ModelVenous SVRRegression ModelArterial SVRCutoff value0.3520.6870.2890.6690.1000.687Sensitivity(95% CI)0.900(0.750-1.000)0.950(0.842-1.000)1.000(1.000–1.000)1.000(1.000–1.000)1.000(1.000–1.000)1.000(1.000–1.000)Specificity(95% CI)0.800(0.500-1.000)0.800(0.500-1.000)0.615(0.333–0.875)0.615(0.333–0.889)0.539(0.250–0.800)0.692(0.417–0.923)PPV(95% CI)0.900(0.750-1.000)0.905(0.760-1.000)0.583(0.300-0.857)0.583(0.300-0.867)0.739(0.565–0.909)0.810(0.632–0.957)NPV(95% CI)0.800(0.500-1.000)0.890(0.625-1.000)1.000(1.000–1.000)1.000(1.000–1.000)1.000(1.000–1.000)1.000(1.000–1.000)Accuracy(95% CI)0.867(0.733–0.967)0.900(0.800-1.000)0.750(0.550–0.900)0.750(0.550–0.950)0.800(0.667–0.933)0.867(0.733–0.967)AUC(95% CI)0.890(0.736-1.000)0.895(0.731-1.000)0.846(0.639–0.990)0.868(0.672-1.000)0.824(0.644–0.960)0.887(0.720-1.000)PH pulmonary hypertension, PH-LHD pulmonary hypertension associated with left heart disease, SVR surface volume ratio, PPV positive predictive value, NPV negative predictive value The predictive performance of SVR and regression models in validation set Sensitivity (95% CI) 0.900 (0.750-1.000) 0.950 (0.842-1.000) 1.000 (1.000–1.000) 1.000 (1.000–1.000) 1.000 (1.000–1.000) 1.000 (1.000–1.000) Specificity (95% CI) 0.800 (0.500-1.000) 0.800 (0.500-1.000) 0.615 (0.333–0.875) 0.615 (0.333–0.889) 0.539 (0.250–0.800) 0.692 (0.417–0.923) PPV (95% CI) 0.900 (0.750-1.000) 0.905 (0.760-1.000) 0.583 (0.300-0.857) 0.583 (0.300-0.867) 0.739 (0.565–0.909) 0.810 (0.632–0.957) NPV (95% CI) 0.800 (0.500-1.000) 0.890 (0.625-1.000) 1.000 (1.000–1.000) 1.000 (1.000–1.000) 1.000 (1.000–1.000) 1.000 (1.000–1.000) Accuracy (95% CI) 0.867 (0.733–0.967) 0.900 (0.800-1.000) 0.750 (0.550–0.900) 0.750 (0.550–0.950) 0.800 (0.667–0.933) 0.867 (0.733–0.967) AUC (95% CI) 0.890 (0.736-1.000) 0.895 (0.731-1.000) 0.846 (0.639–0.990) 0.868 (0.672-1.000) 0.824 (0.644–0.960) 0.887 (0.720-1.000) PH pulmonary hypertension, PH-LHD pulmonary hypertension associated with left heart disease, SVR surface volume ratio, PPV positive predictive value, NPV negative predictive value Discussion This study identified pulmonary vascular SVR on CT as a key radiomics biomarker for detecting PH and distinguishing PAH from PH-LHD. Venous SVR specifically reflected venous remodeling in the PH-LHD rat model before hemodynamic changes emerged, providing preclinical evidence for its potential utility in identifying early-stage disease (Graphical Abstract). Based on the 2022 ESC/ERS risk stratification criteria [1], a higher proportion of intermediate- and high-risk patients was observed in the PH-LHD group compared to the PAH group in our study. These PH-LHD patients also exhibited higher mPAP and worse cardiac function, which may be associated with their older age and the fact that patients who undergo CTA often present with more apparent clinical symptoms, such as chest tightness. Besides, most PH-LHD patient data were collected during initial assessment for newly diagnosed or acutely decompensated heart failure. Moreover, we observed left atrial enlargement in some subjects in both the control and PAH groups, which may be associated with underlying comorbidities such as chronic hypertension. Current non-invasive CT-based PH prediction studies predominantly rely on multimodal models that integrate additional imaging modalities. While these approaches demonstrate good predictive performance, they face limitations in terms of generalizability and lack the efficiency required for rapid disease screening [28]. Based on this foundation, we focused specifically on CT imaging and discovered that the SVR feature, when reconstructed and extracted automatically, maintains high predictive accuracy even as an independent parameter. SVR serves as an integrative biomarker reflecting vascular structural complexity and biomechanical properties, proven to indicate pulmonary capillary function in lung diseases [29, 30]. In this study, changes in vascular SVR primarily resulted from volume alterations due to vascular dilation [31]. While at the cellular level, a higher SVR may enhance exchange efficiency by increasing surface area, and is implicated in cellular responses to mechanical stress [32, 33]. Arterial and venous SVR show complementary diagnostic profiles. Arterial SVR facilitates initial PH screening, while venous SVR enables further identification of PH-LHD. Venous SVR can also identify PH-LHD even without excluding controls, thereby optimizing diagnostic efficiency [34]. Therefore, our findings also hold particular clinical significance for the management of PAH. The early and accurate diagnosis of PAH remains challenging but is critical for initiating timely targeted therapy and improving prognosis [35]. The combined application of arterial and venous SVR, by identifying PH and excluding PH-LHD, provides a diagnostic pathway for PAH. This readily available CT biomarker could serve as a screening tool in the diagnostic workup of suspected PAH, potentially aiding the decision to proceed with invasive RHC. Moreover, our previous research demonstrated that imaging parameters derived from CTA could assist in diagnosing group 3 PH [3]. In this study, we used an AI platform to achieve automated early disease identification without contrast agents. This operates independently of manual intervention, offering higher efficiency. Furthermore, we identified a subgroup-specific venous morphological phenotype in PH-LHD characterized by a reduction in distal and small veins, as well as increased tortuosity, decreased curvature and decreased torsion, a pattern absent in PAH [36]. This likely reflects compensatory dilation and adaptive remodeling in response to chronic venous hypertension and abnormal shear stress [37], while impaired torsion suggests compromised biomechanics and increased right ventricular load [38]. These venous patterns show promise as differential diagnostic biomarker [39]. Therapeutically, targeting venous remodeling could open up new avenues for treatment, although current strategies remain limited [11]. This study has several limitations, which include the cross-sectional design and cohort size. Since the stringent inclusion criteria requiring both CTA and RHC contributed to the limited sample size, a formal power analysis was not performed. Moreover, a proportion of PH patients were referrals without raw DICOM imaging data and were therefore excluded, which may introduce selection bias toward patients with more complete imaging workups. The small sample size relative to the large number of features may increase overfitting risk despite LASSO and external validation. These exploratory findings require large-scale validation in PH referral populations before clinical translation. Additionally, although multi-center transfer learning improved cross-center segmentation robustness, the use of CT scans from two centers with different protocols without voxel resampling or harmonization prior to feature extraction may have affected radiomic reproducibility [40]. Furthermore, our PH-LHD cohort consisted predominantly of patients with heart failure with preserved ejection fraction (HFpEF), as reflected by the preserved LVEF [41]. Given the established prognostic and hemodynamic differences between HFpEF and heart failure with reduced ejection fraction (HFrEF) patients, the characteristics of pulmonary vascular alterations in HFrEF patients remain to be elucidated [42]. Besides, the elevated PVR observed in the PH-LHD cohort suggested a predominance of combined post-capillary pulmonary hypertension (CpcPH), a subtype associated with a poorer prognosis compared to isolated post-capillary pulmonary hypertension (IpcPH) [43]. Therefore, our findings primarily apply to HFpEF-associated PH-LHD patients with CpcPH, while the behavior of SVR in IpcPH or HFrEF remains uncertain. This underscores the need to investigate pulmonary vascular features across LHD subtypes, in addition to detecting early-stage PH-LHD patients. Moreover, while our preclinical data demonstrate that venous SVR can reflect early venous remodeling before hemodynamic changes in a rat model, whether a similar temporal relationship exists in human PH-LHD and whether SVR can enable truly early clinical diagnosis remains to be validated in prospective human studies. Moreover, due to the lack of precise left atrial volume parameters, the difference in diagnostic performance between pulmonary venous SVR and LAV requires further investigation. Besides, future prospective studies focusing on patients with suspected PAH and other PH subtypes are also warranted to validate the utility of SVR in facilitating earlier diagnosis in this high-risk population. Conclusion Pulmonary vascular SVR derived from non-contrast CT serves as a sensitive biomarker for PH identification and PH-LHD discrimination. Arterial SVR aids PH detection, while venous SVR discriminates PH-LHD from PAH and reflects early remodeling in rat models. Combined with morphological features, SVR enables efficient, non-invasive screening. These findings are exploratory and require validation in larger, well-phenotyped PH cohorts before clinical translation. Supplementary Information Supplementary Material 1.