Overview

Schema Version

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

Name

Swin-UNETR for fetal rs-fMRI brain extraction

Link

https://github.com/NicoloPecco/Swin-Functional-Fetal-Brain-Segmentation

Indexing

Keywords: Transformers, CNN, Medical Imaging Segmentation, MRI, Dataset Size, Input Size, Transfer Learning, fetal brain, resting-state fMRI

Content: MR, NR, OB

RadLex: RID10317, RID12775

Author(s)

Nicolò Pecco

Pasquale Anthony Della Rosa

Matteo Canini

Gianluca Nocera

Paola Scifo

Paolo Ivo Cavoretto

Massimo Candiani

Andrea Falini

Antonella Castellano

Cristina Baldoli

Organization(s)

IRCCS Ospedale San Raffaele

Vita-Salute San Raffaele University

Version

1.0

Contact

ti.rsh@elauqsap.asoralled

Funding

Supported by the Italian Ministry of Health’s Ricerca Finalizzata 2016 grant (RF-2016-02364081).

Ethical review

Conducted in accordance with the Declaration of Helsinki and approved by the Ospedale San Raffaele Ethics Committee (registration no. EK Nr.2174/2016). All participants provided written informed consent.

Date

Published: 2024-06-26

References

[1] Pecco N, Della Rosa PA, Canini M, et al.. "Optimizing Performance of Transformer-based Models for Fetal Brain MR Image Segmentation". Radiology: Artificial Intelligence. 2024 Nov;6(6):e230229.. 2024. doi:10.1148/ryai.230229. PMID: 38922031. PMCID: PMC11605146.

Model

Architecture

Transformer-based Swin-UNETR (hierarchical Swin Transformer with U-Net-like encoder-decoder) implemented in PyTorch/MONAI.

Availability

Code available: https://github.com/NicoloPecco/Swin-Functional-Fetal-Brain-Segmentation (original Swin-UNETR: https://github.com/Project-MONAI/research-contributions/tree/main/SwinUNETR)

Clinical benefit

Automated fetal brain extraction from rs-fMRI to facilitate preprocessing and analysis.

Clinical workflow phase

Image preprocessing for fetal rs-fMRI data analysis.

Degree of automation

Fully automated segmentation.

Indications for use

Segmentation (brain extraction) of the fetal brain on resting-state functional MRI volumes from fetuses approximately 21–37 gestational weeks in research settings; applicable across 1.5-T and 3-T scanners in the datasets studied.

Input

Resting-state fetal MRI volumes (rs-fMRI); datasets from 1.5-T and 3-T scanners.

Limitations

Lower representation of earlier gestational weeks (GW < 25) in the internal dataset; relatively small sample size; external validation limited by availability of open-source fetal rs-fMRI; transformer pretraining weights (from CT) did not confer advantages; reduced performance during midfetal period compared with late-fetal period.

Output

CDEs: RDE1759

Description: Binary segmentation mask delineating the fetal brain from surrounding maternal tissues in rs-fMRI volumes.

Reproducibility

Internal 90/10 train-test split with validation; progressive training subset sizes (100%, 66%, 33%); cross-dataset generalization assessed on an external dataset; cross-validation by gestational week; code and internal dataset repository links provided.

Use

Intended: Image segmentation

User

Intended: Radiologist, Researcher