Related Experiment Video
Updated: Oct 15, 2025
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
White matter hyperintensities segmentation using an ensemble of neural networks
Xinxin Li1,2, Yu Zhao1, Jiyang Jiang3
1Key Laboratory of Biomechanics and Mechanobiology, Ministry of Education, Beijing Advanced Innovation Center for Biomedical Engineering, School of Biological Science and Medical Engineering, Beihang University, Beijing, China.
This study introduces an advanced AI pipeline for automatically detecting and measuring white matter hyperintensities (WMHs), a key marker of small vessel disease. The new method significantly improves segmentation accuracy compared to existing tools.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Image Analysis
Background:
- White matter hyperintensities (WMHs) are common indicators of cerebral small vessel disease (CSVD).
- Accurate segmentation and volume estimation of WMHs are crucial for clinical assessment.
- Current automated methods for WMH analysis have limitations in performance and generalizability.
Purpose of the Study:
- To develop and evaluate a novel deep learning pipeline for automated WMH segmentation and volume estimation.
- To compare the proposed pipeline's performance against existing methods using diverse datasets.
- To assess the model's generalization capabilities across different data sources.
Main Methods:
- A deep fully convolutional network ensemble model combining U-Net, SE-Net, and multi-scale features was developed.
- The pipeline was trained and validated on clinical routine (CNSR) and research (MWC) datasets.
- Performance was benchmarked against LGA, LPA, UBO detector, and U-Net using Dice Similarity Coefficient (DSC) and other metrics.
Main Results:
- The proposed pipeline achieved superior performance on both clinical (DSC = .783) and research (DSC = .833) datasets, significantly outperforming other methods (p < .001).
- The model trained on research data demonstrated better generalization (DSC = .736) than that trained on clinical data (DSC = .622).
- The system provides automated whole brain, lobar, and anatomical WMH labeling with both image and text outputs.
Conclusions:
- The developed deep learning pipeline offers a highly accurate and robust solution for automated WMH segmentation.
- This method surpasses current widely used pipelines in WMH analysis.
- The publicly available software and models facilitate broader adoption and further research in CSVD neuroimaging.
