A 3D Deep Learning Model for Assessing Concomitant Pulmonary Embolism Risk by Integrating Clinical and Computed Tomography Venography Features in Patients with Lower Extremity Deep Venous Thrombosis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China; Department of Nuclear Medicine, Chongqing University Affiliated Renji Hospital (The Fifth People's Hospital of Chongqing), Chongqing, China.
- Department of Radiology, Second Affiliated Hospital of Army Medical University (Xinqiao Hospital), Chongqing, China.
- Institute of Research, Infervision Medical Technology Co., Ltd, Beijing, China.
- Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China; Department of Radiology, Changdu People's Hospital of Xizang, Xizang, China. Electronic address: [email protected].
Abstract
The aim of this study was to develop and validate a multimodal deep learning (DL) model integrating clinical and computed tomography venography (CTV) features for predicting concomitant acute pulmonary embolism (PE) risk in patients with lower extremity deep venous thrombosis (LEDVT). This multicentre retrospective study enrolled patients with LEDVT undergoing both computed tomography pulmonary angiography and lower extremity CTV. Centre 1 data were used for model training and internal validation, while data from centres 2 and 3 were used for external testing. Clinical models (logistic regression [LR], random forest [RF], support vector machine [SVM], and XGBoost), 3D DL models (VGG-11, VGG-16, ResNet-18, Res2Net-50, and DenseNet-121), and combined models (LR, RF, SVM, and XGBoost) integrating clinical, imaging, and DL features were developed to select the best performer, which was further benchmarked against validated clinical risk scores (Wells score and modified Geneva score). A total of 348 patients from centre 1 were randomly split into training and internal validation sets, and 127 patients from two independent institutions formed the external testing cohort. The optimal LR clinical, 3D VGG-16 DL, and SVM combined models achieved external testing area under the curve (AUC) values of 0.776, 0.789, and 0.836, respectively. DeLong tests confirmed that the SVM combined model significantly outperformed the LR clinical model and validated clinical scores (all p < .050) on the external testing set, while yielding a comparable AUC to the 3D VGG-16 DL model. Decision curve analysis indicated higher net benefit for the SVM combined model, alongside superior calibration, and positive net reclassification improvement and integrated discrimination improvement values. The SVM combined model integrating clinical, imaging, and DL features demonstrated robust performance, offering a promising tool for predicting concomitant acute PE in patients with LEDVT.