Challenges and artificial intelligence solutions for clinically optimal hepatic venous vessel segmentation

: Liver vessel identification is crucial for clinical disease assessment and treatment planning, especially concerning local treatment of liver tumors. As artificial intelligence (AI) develops in radiology, opportunities arise to craft models adept at hepatic venous vessel segmentation, opening possibilities for creating patient-specific models of the liver anatomy quickly, despite the diverse features of CT images encountered in clinical settings. This research evaluates the performance of AI models combined with various pre-processing filters for liver vessel segmentation, emphasizing clinically relevant results. A novel evaluation method was introduced to offer more anatomically accurate assessments, moving beyond traditional metrics like the Dice score. Using open-source and proprietary datasets, we implemented residual UNet and Dense UNet in combination with smoothness and vesselness filters. We used a clinical evaluation approach focused on major and minor liver vessels, thereby underscoring the precision of AI outcomes. The Dense UNet model with a specific pre-processing filter produced an average Dice score of 0.8144 in our internal dataset. For the public test dataset, the score was 0.7859. Both scores were higher than those not using pre-processing filters, 0.8052 and 0.7765. Clinical assessments showed 85% of AI predictions accurately identified all wanted vessel structures, though segmentation beyond the vessel borders did occur in half the predictions. This study highlights the effectiveness of AI in liver vessel segmentation, with the Dense UNet model combined with pre-processing filters showing high Dice scores and clinical accuracy. • Introduced a clinical evaluation method for hepatic vessels. • Smoothing and vesselness filter applied with AI networks. • Dense UNet produced mean Dice of 0.8144 in the internal dataset. • Evaluation predicts 85% vessel tree segmented by AI.