Developing a deep learning-based imaging diagnostic framework, PVDNet, for differentiating pulmonary artery sarcoma...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Differentiating pulmonary artery sarcoma (PAS) from pulmonary thromboembolism (PTE) using CT pulmonary angiography (CTPA) is challenging.
• Deep learning (DL) offers potential solutions for improving diagnostic accuracy.

Purpose of the Study:

• To develop and validate a DL-based model, PVDNet, for differentiating PAS from PTE on CTPA.
• To assess the diagnostic performance of PVDNet compared to radiologists.

Main Methods:

• Retrospective analysis of CTPA datasets from 952 patients (PAS, acute PTE [APE], chronic PTE [CPE]) from the CHOICE study.
• Development and internal/external validation of the PVDNet DL framework.
• Comparison of PVDNet performance with four radiologists of varying expertise.

Main Results:

• PVDNet achieved high AUC values in the internal test set (0.972 for PAS, 0.902 for APE, 0.900 for CPE).
• In external validation, PVDNet showed comparable AUC to a senior radiologist (0.973 vs. 0.943).
• PVDNet demonstrated moderate agreement with the senior radiologist (kappa = 0.651).

Conclusions:

• The PVDNet model effectively differentiates PAS from PTE, nearing senior radiologist proficiency.
• Further optimization is needed for PVDNet's performance in distinguishing APE from CPE.