Fully automated portal venous system segmentation from 4D flow MRI using 3D nnU-Net

<h2>Abstract</h2><h3>Background</h3> To develop and validate a unified 3D nnU-Net–based automated segmentation system for portal venous 4D Flow MRI that segments the entire portal vein (PV) system and automatically extracts the main portal vein (MPV), reducing post‐processing time while ensuring reliable flow assessments. <h3>Methods</h3> This study was approved by the institutional review board of our hospital. In this retrospective study, 232 candidates underwent portal venous 4D Flow MRI between January 2020 and November 2024, with 145 participants eligible for analysis. A 3D nnU-Net ResEnc M model was trained on 112 cases (after excluding 2 synthesis-failure cases) and tested on 31 internal cases. Manual segmentations by an experienced radiologist served as the reference standard, with additional comparisons made against a second radiologist. Morphological agreement was assessed using Dice scores for both the PV system and MPV. Flow parameters at the MPV, including flow rate, mean velocity, wall shear stress, oscillatory shear index, and energy loss, were compared using Mann–Whitney U tests, Spearman correlations, Bland–Altman analyses, and intraclass correlation coefficients. <h3>Results</h3> The automated system achieved median Dice scores of 0.81 for the PV system and 0.87 for the MPV relative to the reference radiologist. All MPV flow parameters demonstrated strong correlations, with flow rate showing the highest agreement (ρ ≥ 0.96) and minimal bias. Notably, the AI-based segmentation was completed in under 10seconds, compared with a median of 35minutes for manual segmentation. <h3>Conclusion</h3> The proposed 3D nnU-Net–based segmentation system delivers near–radiologist-level accuracy with dramatically reduced processing time, supporting its clinical utility in portal venous 4D Flow MRI.