Overview

Schema Version

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

Name

Deep Learning Segmentation of Ascites on Abdominal CT Scans for Automatic Volume Quantification

Link

https://pubs.rsna.org/doi/10.1148/ryai.230601

Indexing

Keywords: Ascites, CT, Deep learning, Segmentation, Volume quantification, Cirrhosis, Ovarian cancer

Content: CT, GI, OI

RadLex: RID33263, RID29522

SNOMED: 389026000, 363443007

Author(s)

Benjamin Hou

Sungwon Lee

Jung-Min Lee

Christopher Koh

Jing Xiao

Perry J. Pickhardt

Ronald M. Summers

Organization(s)

National Institutes of Health Clinical Center, Department of Radiology and Imaging Sciences

The Catholic University of Korea, Seoul St. Mary’s Hospital, Department of Radiology

National Cancer Institute, Women’s Malignancies Branch

National Institute of Diabetes and Digestive and Kidney Diseases, Liver Diseases Branch

Ping An Technology, Shenzhen, China

University of Wisconsin School of Medicine & Public Health, Department of Radiology

Version

1.0

License

URL: https://pubs.rsna.org/journal/ai

Funding

Intramural Research Program (1Z01 CL040004) of the NIH Clinical Center; utilized NIH Biowulf high-performance computing resources.

Ethical review

Retrospective study; IRB approval obtained; informed consent waived. HIPAA-compliant for nonpublic datasets.

Date

Published: 2024-06-20

Created: 2023-12-15

References

[1] Hou B, Lee S, Lee J-M, Koh C, Xiao J, Pickhardt PJ, Summers RM. "Deep Learning Segmentation of Ascites on Abdominal CT Scans for Automatic Volume Quantification". Radiology: Artificial Intelligence. 2024;6(5):e230601. 2024-06-20. doi:10.1148/ryai.230601. PMID: 38900043. PMCID: PMC11449171.

Model

Architecture

nnU-Net framework (self-configuring); 3D residual U-Net core; loss = binary cross-entropy + soft Dice (equal weighting); SGD optimizer (initial LR 1e-2), batch size 2.

Availability

Model and curated training annotations to be released at https://github.com/rsummers11/Ascites

Clinical benefit

Automated segmentation and quantification of ascites volume to aid assessment and monitoring in patients with liver cirrhosis and ovarian cancer; concordant with expert measurements.

Clinical workflow phase

Clinical decision support systems; workflow optimization (screening and measurement assistance).

Decision threshold

Ascites detection defined as segmented volume ≥50 mL; <50 mL categorized as no ascites for detection analysis.

Degree of automation

Fully automated segmentation and initial detection; supports human review for outliers and final interpretation.

Indications for use

Segmentation and volumetric quantification of ascitic fluid on abdominopelvic CT in adults, including patients with liver cirrhosis and ovarian cancer, in radiology reading environments.

Input

Abdominopelvic CT volumes (with or without IV contrast).

Instructions

Process CT volumes with nnU-Net default CT preprocessing (clip 0.5th–99.5th percentiles; global z-score normalization). Apply trained 3D model to generate ascites mask; compute volume from voxel spacing. Flag cases with predicted volume ≥50 mL as positive for ascites.

Limitations

Training primarily on TCGA-OV (female, ovarian cancer) may limit generalizability. Reduced performance with 5-mm section thickness; challenges in uncommon cases such as loculated ascites and mesenteric edema. Detection analysis excludes false negatives below 50 mL and any volumes <50 mL; corner cases like hepatic hydrothorax, hemoperitoneum, trauma not evaluated.

Output

CDEs: RDE2501, RDE1955

Description: Voxel-wise segmentation mask of ascites on CT; derived ascites volume (mL/L) and detection flag based on 50 mL threshold.

Recommendation

Use as an automated tool to screen for and quantify ascites on CT with radiologist oversight, especially in cirrhosis and ovarian cancer cohorts; review challenging cases and outliers.

Reproducibility

Five-fold cross-validation with ensemble of top models; report mean±SD and 95% CI across folds; uncertainty estimated via softmax confidence.

Sustainability

Training converged in ~2 days on NVIDIA DGX-1 with A100 GPUs; runtime/energy not reported.

Use

Intended: Detection, Image segmentation

Out-of-scope: Diagnosis

Excluded: Decision support

User

Intended: Radiologist, Researcher

Out-of-scope: Patient

Excluded: Layperson