A generative adversarial network model for improved three-dimensional mapping of pulmonary arteries and veins from...

Preoperative three-dimensional (3D) vascular reconstruction from non-contrast computed tomography (NCCT) is often hindered by limited vessel differentiation, whereas contrast-enhanced computed tomography (CECT) is contraindicated for certain patients. This study examined the use of a generative adversarial network (GAN) for synthetic CECT (Syn-CECT)-like images from NCCT, preliminarily evaluating its clinical potential to assist in perinodular vascular segmentation and 3D reconstruction.

This retrospective study included 212 patients. To mitigate the spatial misalignments caused by respiratory motion, the training phase (130 patients) included paired virtual non-contrast (VNC) and true CECT images derived from spectral computed tomography (CT). A Pix2pixGAN framework integrating an attention mechanism and a 2.5-dimensional (2.5D) strategy was applied to synthesize Syn-CECT. The model was then validated on external test sets (82 patients) from two different CT vendors. A quantitative evaluation was conducted via objective image quality metrics. The vascular reconstruction performance within a 50-mm spherical region around pulmonary nodules was quantitatively assessed according to the local Dice similarity coefficient (LocalDice), fragmentation index (FragIndex), and 95th percentile Hausdorff distance (HD95), alongside subjective physician scoring.

In external testing, Syn-CECT showed favorable structural compared to original NCCT as indicated by the multiscale structural similarity index measurement. For the downstream 3D vascular segmentation task, Syn-CECT yielded measurable improvements over NCCT, reflected by a moderately higher LocalDice (0.89-0.90 vs. 0.85-0.89), a lower FragIndex (3.82-5.92 vs. 4.54-8.68), and a lower HD95 (0.08 vs. 0.09-0.12) (all P values <0.01), suggesting better structural continuity and boundary alignment. In terms of subjective evaluation, Syn-CECT received more favorable median vessel visibility scores (score of 4) than did NCCT (score of 3), offering improved visual clarity for evaluating local vasculature.

Using a GAN model to synthesize CECT-like images from standard NCCT scans can moderately improve the continuity and accuracy of 3D pulmonary vascular reconstructions. This image translation strategy demonstrated preliminary clinical value as an adjunctive tool for sublobar resection planning and may serve as a practical alternative for obtaining enhanced vascular anatomical details, especially for patients with contraindications to iodinated contrast media.