Background/purpose: Current conventional algorithms used for 3-dimensional simulation in virtual hepatectomy still have difficulties distinguishing the portal vein (PV) and hepatic vein (HV). The accuracy of these algorithms was compared with a new deep-learning based algorithm (DLA) using artificial intelligence.
Methods: A total of 110 living liver donor candidates until 2017, and 46 donor candidates until 2019 were allocated to the training group and validation groups for the DLA, respectively. All PV or HV branches were labeled based on Couinaud's segment classification and the Brisbane 2000 Terminology by hepato-biliary surgeons. Misclassified and missing branches were compared between a conventional tracking-based algorithm (TA) and DLA in the validation group.
Results: The sensitivity, specificity, and Dice coefficient for the PV were 0.58, 0.98, and 0.69 using the TA; and 0.84, 0.97, and 0.90 using the DLA (P < .001, excluding specificity); and for the HV, 0.81, 087, and 0.83 using the TA; and 0.93, 0.94 and 0.94 using the DLA (P < .001 to P = .001). The DLA exhibited greater accuracy than the TA.
Conclusion: Compared with the TA, artificial intelligence enhanced the accuracy of extraction of the PV and HVs in computed tomography.
Keywords: artificial intelligence; automatic extraction; deep learning; liver vessels; organ vessels.
© 2021 Japanese Society of Hepato-Biliary-Pancreatic Surgery.