A deep learning model for the interpretable identification of pulmonary thromboembolism from computed tomography pulmonary angiography.
Authors
Affiliations (7)
Affiliations (7)
- Department of Pulmonary and Critical Care Medicine, The Third Affiliated Hospital of Nanjing Medical University (Changzhou Second People's Hospital), Changzhou, China.
- Department of Interventional and Vascular Surgery, The Third Affiliated Hospital of Nanjing Medical University (Changzhou Second People's Hospital), Changzhou, China.
- Department of Oncology and Vascular Intervention, Gaochun People's Hospital, Nanjing, China.
- Department of Radiology, The First Affiliated Hospital With Nanjing Medical University (Jiangsu Province Hospital), Nanjing, China.
- Department of Interventional Radiology, Huaian Hospital of Huai'an City (Huaian Cancer Hospital), Huai'an, China.
- Department of Pulmonary and Critical Care Medicine, The Second Affiliated Hospital of Nanjing Medical University, Nanjing, Jiangsu, China.
- Department of Neurosurgery, The Third Affiliated Hospital of Nanjing Medical University (Changzhou Second People's Hospital), Changzhou, China.
Abstract
The rapid identification of pulmonary thromboembolism (PTE) on computed tomography pulmonary angiography (CTPA) is vital but labor-intensive, often leading to diagnostic delays. We aimed to construct and evaluate a YOLOv11 object detection algorithm capable of automatically highlighting intraluminal filling defects to expedite emergency radiological workflows. A retrospective analysis was conducted on CTPA scans from multiple centers. The dataset was divided into a primary internal cohort (<i>n</i> = 1,368) for model derivation and testing, alongside an independent external cohort (<i>n</i> = 98) to assess generalizability. The diagnostic efficacy of the YOLOv11 architecture was quantified using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Additionally, gradient-weighted class activation mapping (Grad-CAM) was applied to map the spatial distribution of the model's focus, ensuring clinical transparency. During internal testing, the proposed framework yielded an AUC of 0.777 [95% confidence interval (CI): 0.765-0.788], corresponding to a sensitivity of 74.53% and a specificity of 64.26%. When applied to the external cohort, the algorithm's discriminative ability remained consistent with an AUC of 0.778 (95% CI: 0.749-0.806). Notably, the external sensitivity reached 86.75% (specificity: 54.46%). Visual assessments via Grad-CAM saliency maps confirmed that the model accurately localized embolic occlusions within the complex pulmonary arterial tree. Utilizing the YOLOv11 architecture for automated CTPA analysis yields a highly sensitive and visually interpretable screening mechanism. This artificial intelligence-assisted approach holds substantial promise for reducing missed diagnoses and accelerating patient triage in acute clinical settings.