Related Experiment Video
Updated: Jul 12, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
47.7K
A federated stroke segmentation to impact limited data institutions
Summary
This study introduces a federated learning approach for stroke segmentation, improving lesion characterization across institutions without requiring local annotated data. This method enhances stroke care support by leveraging diverse datasets for robust, generalized models.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in healthcare
- Neurology
Background:
- Stroke is a leading cause of death, often requiring detailed brain lesion characterization using Diffusion-Weighted Imaging (DWI).
- Current radiological interpretation of DWI is subjective, and existing computational tools lack generalizability due to single-institution training.
- Limited annotated data at some institutions hinders personalized stroke segmentation solutions.
Purpose of the Study:
- To develop the first federated learning approach for stroke segmentation, enabling cross-institutional data collaboration.
- To create robust stroke segmentation models applicable to institutions with limited or no annotated training data.
- To assess the generalization capabilities of state-of-the-art architectures within a federated learning framework for stroke care.
Main Methods:
- Implementation of a federated learning framework to train stroke segmentation models on diverse, multi-institutional DWI data.
- Aggregation of models trained across institutions to create a generalized solution.
- Evaluation of model performance and generalization across different institutional datasets.
Main Results:
- The federated approach successfully trained robust stroke segmentation models by leveraging diverse institutional data.
- The study demonstrated the feasibility of impacting institutions without local data requirements through collaborative learning.
- Analysis revealed challenges and opportunities for state-of-the-art architectures in federated stroke segmentation.
Conclusions:
- Federated learning offers a viable solution for stroke segmentation, overcoming data limitations and enhancing generalizability.
- This collaborative approach supports stroke care by providing reliable lesion characterization tools across various clinical settings.
- Validation of federated solutions is crucial for their successful transfer into clinical practice for stroke management.
