Augmented intelligence to predict V/Q-matched defects based on dynamic lung perfusion scan.

Objectives: Lung ventilation and perfusion (V/Q) scans are essential in the diagnostic algorithm of pulmonary hypertension in order to exclude thromboembolic disease. The ventilation scan helps to identify V/Q-matched defects due to regional hypoxia. However, it is hard to interpret the scan when there is significant amount of aerosol trapped in the bronchi. Under the assumption that respiratory motion implies regional ventilation status, this study aims to use a deep learning model to predict V/Q-matched defects from solely lung perfusion scan.

Methods: We retrospectively collected 666 V/Q scans from patients with either pulmonary hypertension or suspected pulmonary embolism. Among them, we identified 450 scans with at least one perfusion defect for analysis. Two-minute dynamic anterior projection images immediately after tracer injection were fed into the deep learning model with convolutional long short-term memory (ConvLSTM) architecture to predict V/Q-matched defects. Training data augmentation was performed by applying random translations and rotations. Clinical interpretation of whole V/Q scan, including optional SPECT/CT, was used as the reference standard. We performed 10-fold cross-validation, and the prediction accuracy was assessed by receiver operating characteristic (ROC) analysis.

Results: The clinical interpretation documented mismatched defects in 224 scans while the other 226 scans showed mostly matched defects. The proposed ConvLSTM model converges at around 1000 training epochs. In validation datasets, the area under ROC curve was 0.871±0.049. The overall accuracy was 83%. Conclusion: In this study, we proposed a deep learning model to predict V/Q-matched defects based on solely dynamic lung perfusion scan. This technique would help the interpretation when the ventilation scan was interfered by aerosol trapping.

• Download figure
• Open in new tab
• Download powerpoint