A robust automated segmentation method for white matter hyperintensity of vascular-origin
Haoying He1, Jiu Jiang2, Sisi Peng3
1Department of Neurology, Zhongnan Hospital of Wuhan University, 169# East Lake Road, Wuchang District, Wuhan 430071, China.
Neuroimage
|May 19, 2025
Summary
A new transformer-based deep learning method accurately segments white matter hyperintensities (WMH) from MRI scans. This robust approach shows strong generalizability across diverse datasets and imaging conditions, outperforming existing methods for small vessel disease assessment.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Artificial Intelligence in Medicine
Background:
- White matter hyperintensity (WMH) is a key indicator of small vessel disease (SVD), often leading to cognitive impairment.
- Accurate WMH quantification is crucial for diagnosing and monitoring SVD-related disorders.
- Current automated segmentation methods struggle with generalizability across diverse datasets and imaging parameters.
Purpose of the Study:
- To develop and validate a robust deep learning segmentation method for vascular-origin WMH.
- To assess the generalizability of the proposed method across different MRI systems and acquisition settings.
- To compare the performance of the novel method against existing WMH segmentation techniques.
Main Methods:
- A transformer-based deep learning model was developed for automatic WMH segmentation.
- The model utilized both 3D T1 and 3D T2-FLAIR MRI images.
- Training and testing were performed on an initial dataset of 126 participants; external validation involved two independent datasets (170 and 70 subjects) with diverse MRI scanners and field strengths (1.5T, 3T, 5T).
Main Results:
- The proposed method achieved high performance across all datasets, with median Dice coefficients of 0.78 (primary), 0.72 (external 1), and 0.72 (external 2).
- Relative volume errors were 0.15, 0.50, and 0.47, respectively, demonstrating good accuracy.
- True positive rates were high (0.81-0.92) with manageable false positive rates (0.20-0.40), outperforming comparison methods like LGA, LPA, BIANCA, UBO-detector, and TrUE-Net.
Conclusions:
- The developed transformer-based method offers a reliable and robust solution for segmenting vascular-origin WMH.
- The model demonstrates excellent generalizability, performing well on unseen MRI data from various scanners and protocols.
- This approach is suitable for large-scale cohort studies requiring accurate WMH quantification in clinical practice.
