Improved BG-PVS Quantification in Infant Brain MRI Using Anatomy-Informed Pseudo-Labels for Joint BG and PVS
Junghwa Kang1, Dayeon Bak1, Na-Young Shin2,3
1Department of Biomedical Engineering, Hankuk University of Foreign Studies, Yongin-si, Republic of Korea.
Journal of Magnetic Resonance Imaging : JMRI
|March 20, 2026
Summary
This study introduces an automated deep learning method for segmenting basal ganglia (BG) and perivascular spaces (PVS) in infant brain MRIs. The novel approach achieves robust and efficient segmentation, improving glymphatic system research in infants.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Deep Learning
Background:
- Quantifying basal ganglia (BG) perivascular spaces (PVS) is crucial for understanding the infant glymphatic system.
- Accurate PVS quantification in infants presents significant technical challenges.
Purpose of the Study:
- Develop an automated deep learning (DL) method for segmenting BG and BG-PVS in infant brain MRI.
- Utilize an anatomy-informed pseudo-labeling strategy to enhance segmentation accuracy.
Main Methods:
- Retrospective technical development and validation across three infant cohorts (dHCP, BCP, in-house).
- Manual ground-truth labels generated and validated by experienced researchers and a radiologist.
- Performance evaluated using Dice Similarity Coefficient (DSC), recall, positive predictive value, and Hausdorff distance.
Main Results:
- The proposed DL method demonstrated improved accuracy in BG and BG-PVS segmentation (e.g., BG DSC = 0.91 ± 0.03).
- Achieved high agreement with reference measurements for PVS quantification (r = 0.90-0.99, ICC ≥ 0.96).
- Outperformed alternative automated segmentation approaches.
Conclusions:
- The developed method enables robust and annotation-efficient segmentation of BG and BG-PVS in infant MRIs.
- Facilitates advanced research into the glymphatic system in early development.
