An anatomical knowledge-based MRI deep learning pipeline for white matter hyperintensity quantification associated
Li Liang1, Pengzheng Zhou2, Wanxin Lu3
1School of Electronic and Information Engineering, Harbin Institute of Technology at Shenzhen, Shenzhen, Guangdong, China; Peng Cheng Laboratory, Shenzhen, Guangdong, China.
Summary
A new deep learning method, anatomical knowledge-based U-Net (A-U-Net), accurately segments white matter hyperintensities (WMHs) in brain MRIs. This tool aids in understanding WMH burden and its link to cognitive decline in aging populations.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Diagnostics
Background:
- White matter hyperintensities (WMHs) in strategic brain regions are linked to cognitive impairments in Alzheimer's disease (AD).
- Current deep learning methods for WMH segmentation often overlook crucial anatomical knowledge, limiting lesion localization and decision-making accuracy.
- Integrating anatomical information is vital for improving the precision of WMH segmentation and understanding their clinical significance.
Purpose of the Study:
- To develop and evaluate an anatomical knowledge-based deep learning pipeline (A-U-Net) for simultaneous segmentation and localization of WMHs.
- To assess the performance of A-U-Net compared to existing methods, particularly in cohorts with varying WMH burdens.
- To investigate the association between WMH burden, lesion characteristics in specific brain regions, and cognitive function in elderly individuals.
Main Methods:
- Proposed an A-U-Net pipeline integrating handcrafted anatomical spatial features from brain atlases with a U-Net architecture.
- Evaluated the pipeline using manually annotated data from a WMH segmentation challenge.
- Applied the validated pipeline to a separate public database to analyze WMH burden and cognition correlations.
Main Results:
- A-U-Net significantly improved WMH segmentation performance (p < 0.05) compared to methods lacking anatomical knowledge, showing a 14-17% AUC increase in mild WMH cohorts.
- The stage-wise A-U-Net-two-step method achieved top performance (Dice: 0.86, mHD: 3.06 mm), comparable to state-of-the-art.
- WMH accumulation patterns differed between normal aging and cognitive impairment; lesions in strategic regions (frontal/parietal white matter, corpus callosum) correlated significantly with cognition.
Conclusions:
- A-U-Net is a reliable deep learning tool for segmenting and quantifying brain WMHs in elderly populations.
- The method enhances understanding of the relationship between WMH burden and cognitive status.
- A-U-Net holds value for individual cognitive evaluation and advancing AD research.
Keywords:
Anatomical knowledgeCognitive impairmentDeep learningSegmentationWhite matter hyperintensities
