Generating Lung Ventilation Images with Virtual Non-contrast Images from Dual-Energy CT Scans Using Multi-task...

Regional lung ventilation imaging using xenon-enhanced dual-energy CT (Xe-DECT) offers valuable insight into obstructive pulmonary diseases but remains limited in clinical use due to technical and logistical constraints. In this study, a multi-task conditional generative adversarial network (GAN) was developed to generate deep learning-generated ventilation images (DL-Vent) from virtual non-contrast (VNC) images. A total of 269 scans from 177 patients with COPD or asthma-COPD overlap syndrome (ACOS) were used to train, validate, and test the model. The architecture was designed to simultaneously predict ventilation images and emphysema masks using paired inspiratory and expiratory VNC input images. DL-Vent demonstrated strong similarity to measured Xe-DECT ventilation images (Xe-Vent), with dice similarity coefficients of 0.56 for ventilation defects and 0.88 for ventilation regions. The ventilation defect percentages (VDP) of DL-Vent and Xe-Vent showed a high correlation (r_s = 0.82), and both were similarly correlated with pulmonary function test results, including FEV1 (p = 0.71). Radiologists rated DL-Vent images as "fair to good" (mean score 3.9/5) and reliably differentiated defect patterns between COPD and ACOS (Cramer's V = 0.41, p = 0.03). The proposed model provides a promising alternative for functional lung imaging without requiring xenon administration.