Year: 2026
Type: article
Abstract: Accurate estimation of the future liver remnant (FLR) is essential for planning hepatectomy and portal vein embolization (PVE), yet conventional calculation relies on manual segmentation that is time-intensive and dependent on specialized workstations. Thus, LIVSEG was developed as a single-page web application that streams DICOM or NIfTI imaging to a cloud GPU running TotalSegmentator nnU-Net models for automated Couinaud-segment and hepatic-lesion segmentation, with an integrated editor for human review. In a cohort of 60 patients (38 living liver donors, 22 PVE patients; 82 scans), the absolute difference between automated and manual FLR was 3.24 ± 2.69 percentage points (pp) in donors and 4.49 ± 5.14 pp across PVE scans, and the two methods were statistically equivalent within a ± 5 pp margin in all cohorts. Using institutional FLR thresholds, LIVSEG and manual volumetry yielded concordant classifications in 79 of 82 scans (96.3%). Inference was completed in two minutes per case at a GPU cost of $0.014-0.026 per study. LIVSEG allows models to reach clinical workflows while retaining human oversight.
Institution: Vanderbilt University Medical Center
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,690
OpenAlex ID: W7235050708
