Background:Differentiating pulmonary artery sarcoma (PAS) from pulmonary thromboembolism (PTE) based on CT pulmonary angiography (CTPA) is a big challenge, necessitating the incorporation of other methods, such as deep learning (DL). This study aimed to develop and validate a DL-based model, PVDNet, for differentiating PAS and PTE on CTPA.
Methods:This study retrospectively analyzed CTPA image datasets from the prospective CHinese pulmOnary embolism multimodality Imaging artifiCial intelligencE (CHOICE) study to develop and validate a DL model for differentiating PAS from PTE. CTPA image datasets of 952 patients (470 acute PTE [APE], 363 chronic PTE [CPE], and 119 PAS) from 15 hospitals were included. The training set comprised CTPA images from 590 patients, and the internal test set comprised those from 186 patients, all obtained from the same three centers. CTPA images of 176 patients from 12 centers were used for external validation. A DL framework, PVDNet, was employed to perform fine-grained classification. Meanwhile, CTPA images in the external validation set were independently assessed by four radiologists with different levels of expertise. The main outcome measures were area under the curve (AUC) and the consistency test.
Findings:In the internal test set, PVDNet achieved an AUC of 0.972, 0.902, and 0.900 for PAS (95% CI: [0.945, 0.994]), APE (95% CI: [0.855, 0.944]), and CPE (95% CI: [0.852, 0.946]), respectively. Furthermore, PVDNet model demonstrated effective differentiation between PAS and PTE, showing comparable AUC values to a senior radiologist specialized in pulmonary vascular diseases (SRPV) in the external validation set (0.973 vs. 0.943, p = 0.308). The model achieved moderate agreement with SRPV (kappa = 0.651, p < 0.001), which was the highest among four readers.
Interpretation:PVDNet model could differentiate PAS from PTE, with performance approaching the proficiency level of a senior radiologist specializing in pulmonary vascular diseases. PVDNet's performance in distinguishing APE from CPE requires further optimization.
Funding:National Natural Science Foundation of China (No. 82272081), Medical and Health Science and Technology Innovation Project of Chinese Academy of Medical Science (No. 2021-I2M-1-049), and the National Key Research and Development Program of China (No. 2023YFC2507200).