Improve Pulmonary Embolism Diagnosis with Deep Learning Using SPECT Imaging

Abstract

Objectives: Ventilation/Perfusion single-photon emission computed tomography (V/PSPECT) is an important examination technique for diagnosing pulmonary embolism (PE), yet interpreting V/PSPECT remains challenging. This study aims to develop an automatic deep learning pipeline to enhance PE diagnosis on V/PSPECT. Methods: The pipeline comprises four steps: 1) lung mask images segmentation from V/PSPECT images using a 3D U-Net, 2) calculation of ventilation-perfusion mismatch images based on segmented V/PSPECT images, 3) PE diagnosis with a transformer-based model using multi-modal data. 4) Performance comparison between the PE diagnosis model and radiologist, with the model used to assist radiologist in improving diagnostic performance. Results: 482 V/PSPECT examinations were included as the development dataset. Among these, 239 examinations were randomly assigned for developing the lung mask segmentation model. All data in the development dataset were used for training, validating, and testing the PE diagnosis model. The lung mask segmentation model achieved a Dice coefficient of 0.925. The PE diagnosis model derived from data combination V/PSPECT images and lung mask images achieved highest accuracy of 0.938, F1 score of 0.943 and AUC of 0.975(95%CI: 0.920, 0.996). The performance of radiologist improved with the assistance of PE diagnosis model, with accuracy improving from 0.906 to 0.958, AUC improving from 0.912 (95%CI: 0.837, 0.960) to 0.957 (95%CI: 0.895, 0.988). Conclusion: This study introduced a deep learning pipeline for PE diagnosis, accurately segmented lung masks, and diagnosed PE using routine V/PSPECT examinations. The transformer-based PE diagnosis model can assist radiologist in improving performance and shows potential for clinical application.