Overview

Schema Version

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

Name

Attention-based 3D CNN for detection of new MS FLAIR lesions with GAN-based synthetic data augmentation

Link

https://github.com/CCI-Bonn/HD-MS-Lesion-Detection

Indexing

Keywords: Multiple Sclerosis, Synthetic Data Augmentation, Generative Adversarial Network, FLAIR, Longitudinal follow-up, Lesion detection

Content: MR, NR

RadLex: RID28801, RID10312, RID35806, RID50144

Author(s)

Gianluca Brugnara

Chandrakanth Jayachandran Preetha

Katerina Deike

Robert Haase

Thomas Pinetz

Martha Foltyn-Dumitru

Mustafa A. Mahmutoglu

Brigitte Wildemann

Ricarda Diem

Wolfgang Wick

Alexander Radbruch

Martin Bendszus

Hagen Meredig

Aditya Rastogi

Philipp Vollmuth

Organization(s)

Department of Neuroradiology, Heidelberg University Hospital

Division for Computational Neuroimaging, Heidelberg University Hospital

Department of Neurology, Heidelberg University Hospital

Department of Neuroradiology, Bonn University Hospital

Division for Computational Radiology and Clinical AI, Bonn University Hospital

German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany

Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany

Institute for Applied Mathematics, University of Bonn, Germany

Version

1.0

License

URL: https://pubs.rsna.org/doi/10.1148/ryai.230514

Funding

G.B. supported by the Physician Scientist Program of the University of Heidelberg. P.V. supported by an Else Kröner Clinician Scientist Endowed Professorship (Else Kröner Fresenius Foundation; reference: 2022_EKCS.17). Data for synthetic patient generation provided by OASIS-3 (multiple NIH grants listed in Acknowledgments).

Ethical review

Institutional review boards at Heidelberg University Hospital and Bonn University Hospital approved the study and waived the need for informed consent.

Date

Published: 2024-10-16

References

[1] Brugnara G, Jayachandran Preetha C, Deike K, et al.. "Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation for Classification Tasks in Multiple Sclerosis". Radiology: Artificial Intelligence. 2024;6(6):e230514.. 2024-10-16. doi:10.1148/ryai.230514. PMID: 39412405. PMCID: PMC11605143.

Model

Architecture

Custom attention-based three-dimensional convolutional neural network (inspired by non-local neural networks) for classification of new MS lesions from paired baseline and follow-up MRI; training augmented with a 3D conditional GAN (pix2pix-inspired) that synthesizes FLAIR images with controllable lesion burden. Comparative baseline: 3D ResNet-18.

Availability

Open-source code and trained models at https://github.com/CCI-Bonn/HD-MS-Lesion-Detection

Clinical benefit

Automated detection of newly occurring MS lesions on follow-up MRI may improve generalizability across scanners/protocols and assist longitudinal assessment of disease activity.

Clinical workflow phase

Clinical decision support during image interpretation in longitudinal follow-up.

Degree of automation

Fully automated model training and inference for binary classification of new lesion occurrence from paired MRI without prior manual segmentation.

Indications for use

Detection of new FLAIR lesions at MRI during longitudinal follow-up of patients with multiple sclerosis; intended for use on brain MRI pairs (baseline and follow-up).

Input

Paired brain MRI exams (baseline and follow-up), including FLAIR sequences; training additionally used synthetic FLAIR images generated from OASIS-3 T1 MPRAGE with automated tissue/lesion segmentations.

Limitations

Evaluated on a single disease entity and task; only two network architectures assessed; external validation limited to a single external institution; quality and realism of synthetic data depend on the generative model and may propagate biases; larger multicenter studies and broader tasks recommended.

Output

CDEs: RDE626.0, RDE626.1

Description: Binary classification indicating presence or absence of newly occurring MS lesions at the follow-up examination.

Recommendation

Use synthetic data augmentation to mitigate domain shift and class imbalance when training models for longitudinal MS lesion detection; external validation is necessary before deployment.

Regulatory information

Authorization status: Research-use only; no regulatory authorization stated.

Reproducibility

Fivefold cross-validation on internal cohort; external validation on an independent cohort; no sample-specific post hoc optimization; code and models publicly available.

Use

Intended: Detection and diagnosis