Overview

Schema Version

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

Name

Deep learning ensemble for automated estimation of myelin maturation age from pediatric brain MRI

Link

https://github.com/wasserth/MyelinAge

Indexing

Keywords: Pediatrics, MR Imaging, CNS, Brain/Brain Stem, Convolutional Neural Network, Artificial Intelligence, Pediatric Imaging, Myelin Maturation, Brain MRI, Neuroradiology

Content: MR, NR, PD

RadLex: RID15994, RID10312, RID17044, RID34492, RID49531, RID10796

Author(s)

Tugba Akinci D’Antonoli

Ramona-Alexandra Todea

Nora Leu

Alexandre N. Datta

Bram Stieltjes

Friederike Pruefer

Jakob Wasserthal

Organization(s)

Department of Pediatric Radiology, University Children’s Hospital Basel, Switzerland

Department of Pediatric Neurology and Developmental Medicine, University Children’s Hospital Basel, Switzerland

Institute of Radiology and Nuclear Medicine, Cantonal Hospital Basel, Switzerland

Department of Neuroradiology, Clinic of Radiology and Nuclear Medicine, University Hospital Basel, Switzerland

Department of Research and Analysis, Clinic of Radiology and Nuclear Medicine, University Hospital Basel, Switzerland

Version

1.0

Contact

Tugba Akinci D’Antonoli (corresponding author), email: hc.sabinu@ilonotnadicnika.abgut

Funding

Authors declared no funding for this work.

Ethical review

Retrospective study approved by the Ethics Committee Northwest and Central Switzerland (EKNZ BASEC identification no. 2021–01682) with waiver of informed consent.

Date

Published: 2023-07-26

References

[1] Akinci D’Antonoli T, Todea RA, Leu N, Datta AN, Stieltjes B, Pruefer F, Wasserthal J. "Development and Evaluation of Deep Learning Models for Automated Estimation of Myelin Maturation Using Pediatric Brain MRI Scans". Radiology: Artificial Intelligence. 2023 Sep;5(5):e220292.. 2023-07-26. doi:10.1148/ryai.220292. PMID: 37795138. PMCID: PMC10546368.

Model

Architecture

Ensemble of a 3D convolutional neural network (reimplementation of Chen et al. multichannel 3D CNN) and a 2D EfficientNet-b0 (pretrained with Noisy Student). Implemented in PyTorch Lightning; trained on NVIDIA GeForce RTX 3090.

Availability

Code: https://github.com/wasserth/MyelinAge. Dataset: https://zenodo.org/record/8055666

Clinical benefit

Accurately predicts myelin maturation age to support radiologists, potentially reducing time to assessment and interobserver variability.

Clinical workflow phase

Clinical decision support systems; workflow optimization in imaging interpretation.

Degree of automation

Fully automated prediction of myelin maturation age from T1- and T2-weighted MRI, intended to assist radiologist decision making.

Indications for use

Estimation of corresponding myelin maturation age on brain MRI in infants and young children (0–3 years), in radiology departments.

Input

Axial T1-weighted and T2-weighted pediatric brain MRI scans (preprocessed: de-identified, skull stripped, T1 registered to T2, affinely registered to an 18-month reference).

Instructions

Apply the described preprocessing (registration, skull stripping, normalization) and run the provided PyTorch Lightning models as in the public repository.

Limitations

Training data included minor pathologies; no cases with delayed myelin maturation were available; varying acquisition protocols over a decade (same scanner); increased prediction variance with age; cross-validation (non-nested) may overestimate performance; known failure modes include motion artifacts and inaccurate/excessive skull stripping; dHCP external set limited to <2 months leading to under/overestimation trends.

Output

Description: Regression output: predicted myelin maturation age in months.

Recommendation

Use as an assistive tool alongside expert review for assessing myelin maturation on pediatric brain MRI; not integrated with PACS and not a regulated medical device.

Regulatory information

Comment: Research-use code and data; no regulatory clearance; not PACS-integrated.

Reproducibility

Code and full dataset are openly available; training details provided (fivefold cross-validation, hardware, training/inference times).

Sustainability

Training approximately 45 hours total for 30 models on a single RTX 3090 GPU; inference ~1.19 seconds per sample for the ensemble.

Use

Intended: Detection and diagnosis

Out-of-scope: Detection and diagnosis

Excluded: Decision support

User

Intended: General diagnostic radiologist, Subspecialist diagnostic radiologist, Other

Out-of-scope: Patient, Layperson