What question did this study set out to answer?

May 20, 2026

A15-06 AI-Based Quantification of Pulmonary Blood Volumes From CT Pulmonary Angiography Helps Screen for Pulmonary Hypertension

Key Points

This study aims to leverage AI for quantifying pulmonary blood volumes from CT pulmonary angiography to improve screening for pulmonary hypertension.
Utilized a deep learning model (Vascul8™) for automated segmentation of pulmonary arterial and venous blood volumes on CTPA.
Conducted a secondary analysis of data from the Cambridge PH Registry for screening pulmonary hypertension.
Performed logistic regression modeling with a randomized allocation of patients into training and testing datasets.
AI-based PBV model showed AUROC of 0.94 for training and 0.85 for testing, with sensitivity of 98% and specificity of 67%.
Compared to manual vessel diameter measurements, the AI model had better performance in detecting pulmonary hypertension.
AI model also outperformed echocardiography with an AUROC of 0.75, revealing significant strengths at lower mean pulmonary artery pressure thresholds.

Abstract

Abstract Rationale CT pulmonary angiography (CTPA) provides clinicians with information on the probability of pulmonary hypertension (PH) through manual measurements of the main pulmonary artery (PA) diameter and the PA-to-aorta diameter ratio. However, data from CTPA for PH screening are currently underleveraged. This study aims to use currently untapped CTPA data by harnessing AI-based quantification of pulmonary blood volumes (PBV) as the marker of PH. Methods We used a deep learning model (Vascul8™) for automated segmentation and quantification of pulmonary arterial (PA) and venous (PV) blood volumes on CTPA. The proximal PA and PV were separated by the boundaries of the mediastinal pleura and labeled as “central.” A secondary analysis of patients from the prospective Cambridge PH Registry (CAPHTURE) was performed to screen for PH. PH was defined as a mean pulmonary artery pressure (mPAP) of 20 mmHg on right heart catheterization. Patients were randomly allocated (60:40) into training and testing datasets for logistic regression modelling. The performance of the AI-based PBV model was compared with both the manual vessel diameter measurement model (main PA diameter and PA-to-aorta ratio) and echocardiography. Results A total of 376 CTPA scans from 56 UK hospitals were analysed—331 with PH of different aetiologies (median age 60 years, 57% female) and 45 without PH (median age 76 years, 84% female). Both the training and testing of the AI-based PBV model, utilizing central PA and PV blood volumes (AUROC 0.94 and 0.85; sensitivity 98% and 95%; specificity 67% and 50%), outperformed the manual vessel diameter measurement model (AUROC 0.89 and 0.84; sensitivity 97% and 99%; specificity 30% and 28%) in screening for PH. In particular, the AI-based PBV model demonstrated better performance in PH due to lung disease (AUROC 0.93 and 0.90 for training and testing, respectively) compared with the manual vessel diameter model (AUROC 0.86 and 0.79 for training and testing, respectively). The AI-based PBV model also outperformed echocardiography (AUROC 0.75, sensitivity 88%, and specificity 62%) in screening for PH. This was specifically observed for mPAP thresholds of ≤ 30 mmHg (AUROC 0.85 vs. 0.63) and ≤40 mmHg (AUROC 0.86 vs. 0.69). Conclusions Automated AI-based quantification of PBV from CTPA represents a potential tool to non-invasively and rapidly screen for PH at scale. This abstract is funded by: None

Bookmark

Cite This Study

Walsh et al. (Fri,) studied this question.

synapsesocial.com/papers/6a0d4fd2f03e14405aa9b4ee https://doi.org/https://doi.org/10.1093/ajrccm/aamag162.2204

Bookmark