Accuracy of artificial intelligence-based simulation for assessing lung vessels and volume using unenhanced computed tomography.

Objectives: The advantages of preoperative three-dimensional (3D) image simulations, which require enhanced computed tomography (ECT), for anatomical lung resection are well documented. However, the necessity for contrast agent presents a significant barrier for some patients. This study thus aims to evaluate the accuracy of an artificial intelligence-based 3D simulation using unenhanced computed tomography (UECT) data in comparison to ECT data.

Methods: The study enrolled 18 lung cancer patients who underwent anatomical lung resections. Utilizing the artificial intelligence software Version6.7 within the Synapse Vincent system (Fujifilm Corporation, Tokyo, Japan), automatic construction of 3D images of the bronchovascular trees was achieved using both ECT and UECT. We further assessed the accuracy of pulmonary vessel identification on UECT, and compared the calculated lung segment volumes obtained from UECT with those obtained from ECT.

Results: The comparison of accuracy to operative findings showed that ECT identified 98.9% of artery branches (PAs) and 85.7% of vein branches (PVs), while UECT identified 96.6% of PAs and 82.1% of PVs. Out of 371 PAs and 319 PVs identified on ECT, UECT failed to detect 16 PAs (4.4%) and 32 PVs (10.1%), yielding a correlation coefficient for branch detection of 0.9783 (P < 0.001). There was a significant correlation between ECT and UECT in measuring artery-oriented volumes on both the right-side segments (R = 0.8330) and the left-side segments (R = 0.8082).

Conclusions: This 3D image technique using UECT data may be comparable to that obtained with ECT data in terms of achieving lobar and partial segmental branch levels.