Overview

Schema Version

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

Name

DeepAll nnU-Net liver segmentation (MRI-trained)

Link

https://dx.doi.org/10.1148/ryai.220080

Indexing

Keywords: Convolutional Neural Network, Deep Learning, Supervised Learning, CT, MRI, Liver Segmentation, domain shift, generalizability, nnU-Net, multisource training

Content: CT, GI, MR

Author(s)

Brandon Konkel

Jacob Macdonald

Kyle Lafata

Islam H. Zaki

Erol Bozdogan

Mohammad Chaudhry

Yuqi Wang

Gemini Janas

Walter F. Wiggins

Mustafa R. Bashir

Organization(s)

Department of Radiology, Duke University School of Medicine, Duke University Medical Center

Department of Radiation Oncology, Duke University School of Medicine

Department of Medicine, Division of Gastroenterology, Duke University School of Medicine

Department of Electrical & Computer Engineering, Duke University Pratt School of Engineering, Duke University

Department of Radiology, Faculty of Medicine, Benha University, Benha, Egypt

Department of Radiology, College of Medicine-Tucson, University of Arizona, Tucson, AZ

Department of Radiology, Rutgers Health-Newark Beth Israel Medical Center, Newark, NJ

Version

1.0

License

Contact

moc.liamg@leknok.nodnarb

Funding

Authors declared no funding for this work.

Ethical review

HIPAA-compliant, retrospective study approved by the institutional review board with waiver of informed consent.

Date

Updated: 2023-05-01

Published: 2023-02-22

Created: 2022-04-22

References

[1] Konkel B, Macdonald J, Lafata K, Zaki IH, Bozdogan E, Chaudhry M, Wang Y, Janas G, Wiggins WF, Bashir MR. "Systematic Analysis of Common Factors Impacting Deep Learning Model Generalizability in Liver Segmentation". Radiology: Artificial Intelligence. 2023;5(3):e220080.. 2023-02-22. doi:10.1148/ryai.220080. PMID: 37293348. PMCID: PMC10245179.

Model

Architecture

nnU-Net (2D) dynamic, self-configuring U-Net-based fully automatic segmentation framework; fivefold ensemble at inference.

Availability

Models were trained using the publicly available nnU-Net framework; specific trained weights were not reported as publicly available.

Clinical benefit

Automated liver segmentation to facilitate volumetry and selection of intrahepatic voxels for quantitative analysis; explores factors affecting cross-domain generalizability.

Degree of automation

Fully automatic segmentation with postprocessing (largest connected component).

Indications for use

Research-use liver segmentation on abdominal MRI (various T1/T2 sequences) with demonstrated generalization to CT arterial/portal/delayed phases; not evaluated as a clinical device.

Input

Abdominal MRI volumes (T1-weighted dynamic phases, non–fat-suppressed T1, opposed-phase, T2-weighted ssfse); generalization tested on CT phases (arterial, portal, delayed).

Instructions

Preprocessing followed nnU-Net: DICOM to NIfTI; MRI voxel intensities z-score normalized per 2D slice; CT truncated to −250 to 200 HU then normalized similarly. Trained with fivefold cross-validation (1000 epochs/fold), SGD with Nesterov momentum (μ=0.99), initial LR 0.01 with poly decay; loss = cross-entropy + Dice. Data augmentation included rotation/scaling, Gaussian noise/blur, brightness/contrast, low-resolution simulation, gamma augmentation, mirroring. Inference ensembles 5 folds; keep largest connected component.

Limitations

Training data excluded large liver tumors; DeepAll struggled in areas containing large tumors in LiTS. Only 2D nnU-Net used; 3D/ensemble variants not explored for cross-domain effects. Patient-level factors (fat/iron deposition, disease state) not analyzed. Single-source ssfse model generalized poorly to other types. Training used only MRI; effect of multimodal training not assessed.

Output

CDEs: RDE1220, RDE1194

Description: Binary liver segmentation mask per volume (3D) with evaluation by Dice-Sørensen coefficient.

Recommendation

Diversify MRI types in training data to improve cross-vendor, cross-sequence, and cross-modality generalizability for liver segmentation.

Regulatory information

Authorization status: Not a regulated/cleared medical device; research study

Reproducibility

Fivefold cross-validation with ensemble inference; dataset-specific nnU-Net self-configuration reported (details in Table S6).

Use

Intended: Image segmentation

Out-of-scope: Detection, Diagnosis

Excluded: Other

User

Intended: Radiologist, Researcher

Out-of-scope: Patient

Excluded: Layperson