Video-based Deep Learning for Pulmonary Embolism Classification on Paired Contrast-enhanced and Noncontrast Chest CT: A Single-center Retrospective Feasibility Study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Chest Diseases, Faculty of Medicine, Kahramanmaraş Sütçü İmam University, Kahramanmaraş, Turkey (H.İ.Ç. and H.Ş.). Electronic address: [email protected].
- Department of Electrical and Electronics Engineering, Engineering Faculty, Osmaniye Korkut Ata University, Osmaniye, Turkey (K.M.S.). Electronic address: [email protected].
- Department of Electrical and Electronics Engineering, Kahramanmaraş Sütçü İmam University, Kahramanmaraş, Turkey (A.A.); Lodz University of Technology, Institute of Electronics, Lodz, Poland (A.A.). Electronic address: [email protected].
- Department of Chest Diseases, Faculty of Medicine, Kahramanmaraş Sütçü İmam University, Kahramanmaraş, Turkey (H.İ.Ç. and H.Ş.). Electronic address: [email protected].
Abstract
Pulmonary embolism (PE) is a disease that is difficult to diagnose because its clinical findings are nonspecific and it carries high mortality when not treated early. Computed tomography pulmonary angiography (CTPA), the gold standard for diagnosis, cannot be performed in every patient because of contraindications related to contrast medium; Noncontrast computed tomography (NCCT), on the other hand, cannot be used as a reliable tool for exclusion because of its low sensitivity. In this retrospective study, the paired images of 716 patients (129 PE-positive, 587 PE-negative) who had undergone both CTPA and NCCT on the same day between 2018 and 2026 were examined. The availability of two scans obtained from the same patient on the same day provided a paired cohort in which clinical and anatomical variables other than the presence of contrast were largely held constant, so that the CTPA scan established with certainty where a thrombus lay in each patient whose noncontrast scan was analyzed. Each scan was converted into a 12-second mediastinal-window video, from which a 40-frame segment centered on the temporal midpoint of the recording was extracted for analysis. For the automatic detection of thrombus, three-dimensional deep learning architectures have been developed: PureStream-3D, BiLSTM-ResNet-3D based on bidirectional long short-term memory (BiLSTM), and AttnFocus-3D based on an attention mechanism. The models have been optimized with different training configurations to address class imbalance, and the dataset has been divided at the patient level into 70% training, 20% validation, and 10% independent test sets. In the CTPA test set, the highest discrimination has been achieved by the AttnFocus-3D-1 model (accuracy 0.9306; area under the curve [AUC] 0.98), whereas the most balanced sensitivity-specificity profile has been obtained with the BiLSTM-ResNet-3D-2 variant (sensitivity 0.9138, specificity 0.8571). In the NCCT test set, the AttnFocus-3D-2 architecture has reached an accuracy of 0.9306, an AUC of 0.97, a sensitivity of 0.9655, and a specificity of 0.7857. Taken together, these findings indicate that the developed deep learning approaches are strong in confirming the presence or absence of PE and can be considered as a decision-support infrastructure in clinical scenarios with high diagnostic uncertainty.