Automated CT Pulmonary Angiography-Derived Periclot Radiomics for CTEPH Detection

CTEPH remains underdiagnosed and can be difficult to differentiate on routine CTPA from acute PE and from patients without pulmonary thromboembolism. We developed and evaluated a fully automated CTPA-based machine learning algorithm combining vascular blood volume metrics with clot and periclot imaging features. In this retrospective multicenter study, intraparenchymal pulmonary vessels were segmented using a scale-space particle approach, deep learning models estimated vessel radius, clot probability, and clot area from cross-sectional patches along the vessel axis, and subject-level radiomic and blood volume features were used to train an XGBoost classifier on a Northwestern University (NU) cohort (89 CTEPH; 176 controls). The trained model was then applied (without retraining) to held-out NU data, an independent acute PE cohort (152 patients) as an exploratory generalization test, and an external cohort from Brigham and Women's Hospital (23 CTEPH; 34 controls). Among 474 participants (mean age, 51 years ± 16; 299 men), the model achieved an AUC of 0.87 (95% CI: 0.79, 0.94) for CTEPH versus controls in internal testing, 0.74 (95% CI: 0.61-0.86) in external testing, and 0.89 (95% CI 0.83-0.93) for CTEPH versus acute PE. Excluding periclot features reduced discrimination, particularly for CTEPH versus acute PE (AUC 0.71; 95% CI: 0.63, 0.79). These findings support the feasibility of automated CTPA-derived vascular, clot, and periclot features for differentiating CTEPH from acute PE and controls without evidence of pulmonary thromboembolism, with further validation needed to assess generalizability across institutions.