https://atlas.rsna.org/schemas/2025-11/model.json

CT-based Liver Couinaud Segment and Spleen Segmentation for LSVR and Cirrhosis Prediction

https://dx.doi.org/10.1148/ryai.210268

1.0

Supported in part by the Intramural Research Program of the National Institutes of Health, Clinical Center, National Institute of Diabetes and Digestive and Kidney Diseases, and National Cancer Institute.

Multi-institutional retrospective HIPAA-compliant study; IRB approval at each institution with waiver of additional informed consent.

Two in-house deep learning segmentation models (details in supplement); one for liver Couinaud segments (I–VIII) and one for spleen.

Noninvasive, objective quantification (LSVR and spleen volume) to aid diagnosis of cirrhosis and advanced fibrosis on contrast-enhanced CT; performance similar to manual measurements and enables multivariable risk models.

Clinical decision support systems; workflow optimization via automated segmentation and volumetry.

Fully automated segmentation and measurement of liver segments and spleen from CT; automated derivation of LSVR and attenuation statistics.

Assessment of chronic liver disease severity (advanced fibrosis and cirrhosis) in adults undergoing abdominal contrast-enhanced CT (portal venous phase) in radiology practice.

Abdominal contrast-enhanced CT volume (portal venous phase).

Use portal venous phase CT including full liver coverage. Exclude scans without IV contrast, without full liver coverage, or in patients post hepatectomy/splenectomy. Outputs include segmentations, volumes, and attenuation statistics; LSVR is computed as (segments I–III volume)/(segments IV–VIII volume).

External performance varied by etiology and biopsy staging system; lower performance in non-HCV populations. Known failure modes include mis-segmentation of caudate lobe (due to ground truth issues), under-segmentation of left lateral segments when liver wraps around spleen, and occasional spleen under/over-segmentation including adjacent stomach. Training and manual ground truths exhibited inter-reader variability. Heterogeneous scanners/protocols over long accrual period. Needle biopsy sampling and staging differences (METAVIR vs Ishak/Knodell) may introduce label noise.

Use as an explainable adjunct to radiologist assessment to quantify morphologic changes associated with fibrosis/cirrhosis; consider disease etiology when interpreting results.

Automated algorithm yields consistent outputs on the same input; inter-reader variability in manual measurements noted, highlighting improved consistency of automated measurements.