Overview

Schema Version

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

Name

Federated nnU-Net for tumor segmentation (FILTS and FeTS study)

Link

https://dx.doi.org/10.1148/ryai.220082

Indexing

Keywords: Federated deep learning, nnU-Net, Tumor segmentation, CT, MRI, Liver, Brain, Glioma, Earth mover’s distance, Bhattacharyya distance, Chi-square distance, Kolmogorov–Smirnov distance, Data distribution, Non-IID data

Content: CT, MR, GI, NR, OI, IN, RS

RadLex: RID11085, RID39127, RID10782, RID10321, RID6189, RID16585

SNOMED: 25370001, 1156974002, 443936004, 1163375002, 278527001, 1157043006, 85057007, 109841003, 1260050006

Author(s)

Guibo Luo

Tianyu Liu

Jinghui Lu

Xin Chen

Lequan Yu

Jian Wu

Danny Z. Chen

Wenli Cai

Organization(s)

Department of Radiology, Massachusetts General Hospital and Harvard Medical School

Intel Corporation

Department of Statistics and Actuarial Science, The University of Hong Kong

College of Computer Science and Technology, Zhejiang University

Department of Computer Science and Engineering, University of Notre Dame

Version

1.0

License

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

Contact

ude.dravrah.hgm@ilneW.iaC

Funding

Supported by the National Institutes of Health/National Cancer Institute (grant no. R42CA189637) and the Children’s Tumor Foundation (grant no. CTF-2021-10-02).

Ethical review

Retrospective, HIPAA-compliant study approved by the institutional review board; need for patient informed consent was waived.

Date

Updated: 2023-05-01

Published: 2023-04-26

Created: 2022-04-26

References

[1] Luo G, Liu T, Lu J, Chen X, Yu L, Wu J, Chen DZ, Cai W. "Influence of Data Distribution on Federated Learning Performance in Tumor Segmentation". Radiology: Artificial Intelligence. 2023-05-01. doi:10.1148/ryai.220082. PMID: 37293342. PMCID: PMC10245185.

Model

Architecture

Federated implementation of nnU-Net (3D U-Net pipeline) using server–client topology and Fed-Avg aggregation.

Availability

The data and scripts used to perform study evaluations will be made publicly available (per Data Availability statement).

Clinical benefit

Research evaluation of federated deep learning for tumor segmentation; insights into effects of inter-site data distribution differences.

Clinical workflow phase

Research and algorithm development; benchmarking of federated learning on multicenter datasets.

Degree of automation

Automated image segmentation model training and inference within a federated learning framework; decision support not intended for clinical use.

Indications for use

Research-only: segmentation of liver tumors on contrast-enhanced CT and brain tumors (glioma/glioblastoma; enhancing and nonenhancing regions) on postcontrast T1-weighted MRI in a federated learning setting.

Input

CT images of liver (contrast-enhanced; arterial/portal venous/delayed) and postcontrast T1-weighted brain MRI (plus FeTS preprocessing: coregistration, resampling to 1×1×1 mm, skull stripping).

Instructions

Run nnU-Net planning/preprocessing (3D pipeline) uniformly; train federated and centralized models with identical preprocessed data and hyperparameters. Federated setup: one server and two clients (each with one subdataset). 80:20 train/test split within each group. Trained on NVIDIA Tesla P40 GPUs (24 GB).

Limitations

Site A (LiTS) is a multisite dataset; sites B and C are single-site—protocol heterogeneity may affect findings. Tumor type not reported for sites A and B. Potential interreader variability across institutions and software. Results specific to Fed-Avg and server–client topology; not all Fed-DL algorithms assessed.

Output

CDEs: RDE2038, RDE1401, RDE1956, RDE1400, RDE1402, RDE1404

Description: Voxelwise tumor segmentation masks; evaluation reports include Dice and θ (federated-to-centralized Dice ratio).

Recommendation

To achieve federated performance comparable to centralized training, use client datasets with small inter-site distribution distances; consider data augmentation/domain adaptation to reduce distances (e.g., as measured by EMD, BD, CSD).

Regulatory information

Comment: No regulatory clearance; methodological research on federated learning.

Authorization status: Research study; not a regulated medical device.

Reproducibility

Same preprocessing and hyperparameters used for federated and centralized models; random 80/20 splits within each group; θ reported as mean ± SE (SE via bivariate first-order Taylor expansion).

Sustainability

Training performed on NVIDIA Tesla P40 GPUs (24 GB). Computational/energy metrics not reported.

Use

Intended: Image segmentation

Out-of-scope: Diagnosis

Excluded: Decision support

User

Intended: Researcher

Out-of-scope: Patient

Excluded: Subspecialist diagnostic radiologist