NeoLUNet: U-Net ensemble for neonatal lung MRI segmentation
2025-12-03https://doi.org/10.1148/atlas.1764790893933
31
Overview
Schema Version
https://atlas.rsna.org/schemas/2025-11/model.json
Name
NeoLUNet: U-Net ensemble for neonatal lung MRI segmentation
Link
https://github.com/SchubertLab/NeoLUNet
Indexing
Keywords: Bronchopulmonary Dysplasia, Chronic Lung Disease, Preterm Infant, Lung Segmentation, Lung MRI, BPD Severity Assessment, Deep Learning, Lung Imaging Biomarkers, Lung Topology
Content: MR, CH, PD
RadLex: RID35976, RID10312, RID10795, RID28799
Author(s)
Benedikt Mairhörmann
Alejandra Castelblanco
Friederike Häfner
Vanessa Koliogiannis
Lena Haist
Dominik Winter
Andreas Flemmer
Harald Ehrhardt
Sophia Stöcklein
Olaf Dietrich
Kai Förster
Anne Hilgendorff
Benjamin Schubert
Organization(s)
Helmholtz Zentrum München, Computational Health Center
Institute for Lung Health and Immunity and Comprehensive Pneumology Center (DZL)
Institute of AI for Health, Helmholtz Zentrum München
Perinatal Center, Hospital of the Ludwig-Maximilian University, Munich
Department of Radiology, Hospital of the Ludwig-Maximilian University, Munich
Center for Comprehensive Developmental Care (CDeCLMU), Dr. von Hauner Children’s Hospital
Department of General Pediatrics & Neonatology, Justus-Liebig-University (DZL), Giessen
Division of Neonatology and Pediatric Intensive Care Medicine, University Medical Center Ulm
Department of Mathematics, Technical University of Munich
Version
1.0
License
Text: CC BY 4.0
URL: https://creativecommons.org/licenses/by/4.0/
Funding
Young Investigator Grant NWG VH-NG-829 (Helmholtz Foundation/Helmholtz Zentrum München), German Center for Lung Research (DZL; German Ministry of Education and Health, BMBF), DFG Research Training Group GRK2338 (Targets in Toxicology), Stiftung AtemWeg (LSS AIRR), Helmholtz Zentrum München Postdoctoral Fellowship Program, BMBF-funded de.NBI Cloud within the German Network for Bioinformatics Infrastructure.
Ethical review
Prospective study with informed parental consent; ethics approvals cohort 1 EC#195–07 and cohort 2 EC#135–12.
Date
Published: 2023-10-25
References
[1] Mairhörmann B, Castelblanco A, Häfner F, et al.. "Automated MRI Lung Segmentation and 3D Morphologic Features for Quantification of Neonatal Lung Disease". Radiology: Artificial Intelligence. 2023 Nov;5(6):e220239.. 2023-10-25. doi:10.1148/ryai.220239. PMID: 38074782.
Model
Architecture
Ensemble of 2D U-Net convolutional neural networks (trained per rater) with pixelwise majority voting; additional 3D U-Net variants evaluated.
Availability
Source code and resources available at https://github.com/SchubertLab/NeoLUNet; Instant-DL framework adapted for training.
Clinical benefit
Automated, expert-level neonatal lung MRI segmentation enabling standardized quantification of lung volume and 3D morphologic features; supports radiation-free assessment and stratification of bronchopulmonary dysplasia severity.
Clinical workflow phase
Clinical decision support systems; imaging quantification to complement diagnostic assessment.
Degree of automation
Fully automated segmentation and feature extraction; models for classification provide decision support.
Indications for use
Research-use automated MRI analysis in preterm neonates near-term age, with or without bronchopulmonary dysplasia, to segment lungs and compute 3D morphologic features for disease assessment in hospital MRI settings.
Input
Axial T2-weighted HASTE neonatal lung MRI at 3T; images cropped to 128×128 pixels for 2D U-Net training.
Instructions
Acquire quiet-breathing axial T2-weighted HASTE neonatal lung MRI at 3T with approximately 1.3×1.9 mm in-plane resolution, 4 mm slice thickness, and 0.4 mm gap. Apply the U-Net ensemble with majority voting to generate lung masks; reconstruct 3D lung volumes and compute predefined morphologic features.
Limitations
Performance decreases with lower image quality; external cohort showed slightly lower VDC (0.880) versus internal cross-validation (0.908). Training and validation performed on two centers with T2-weighted acquisitions only; generalizability to other protocols, scanners, or pathologies not assessed. 3D U-Net underperformed on current dataset. Not a regulated clinical device.
Output
CDEs: RDE1878, RDE2471, RDE1213, RDE1708, RDE1201, RDE289
Description: Binary lung segmentation masks enabling 3D reconstruction and automated calculation of 78 morphologic features (volumetric, intensity, surface) per lung; downstream models output BPD severity classification probabilities and estimates of respiratory support duration.
Recommendation
Use as a standardized, automated MRI analysis tool to complement clinical assessment of neonatal lung disease and BPD severity in research settings.
Regulatory information
Comment: Academic research model; no formal regulatory submission reported.
Authorization status: Not a regulated/cleared medical device; research study.
Reproducibility
Leave-one-patient-out cross-validation in cohort 1 and independent validation in cohort 2; code repository provided for reproducibility.
Use
Intended: Prognosis, Image segmentation, Risk assessment
Out-of-scope: Other
Excluded: Decision support
User
Intended: Radiologist, Subspecialist diagnostic radiologist, Researcher
Out-of-scope: Patient, Layperson
Excluded: Other