Abstract
Alzheimers Dement. 2025 Dec;21 Suppl 2:e104470. doi: 10.1002/alz70856_104470.
ABSTRACT
BACKGROUND: Enlarged perivascular spaces (PVS) are imaging biomarkers of cerebral small vessel disease (CSVD) associated with age, hypertension, and neurodegenerative conditions. Despite their clinical relevance, accurate quantification of PVS on T1-weighted magnetic resonance imaging (MRI) remains challenging due to both variability across imaging protocols, and their small size and limited contrast. Automated methods such as convolutional neural networks (CNNs) offer a scalable solution, but existing tools are limited in performance and generalizability.
METHOD: This study introduces segcsvdPVS, a CNN-based tool designed for automated PVS segmentation on T1-weighted images, based on a hierarchical framework incorporating anatomical information and robust training strategies. It was trained on semi-automated RORPO-based ground truth data and validated using both manual and semi-automated segmentations. A large and comprehensive cohort (n = 1351) spanning multiple datasets characterized by diverse imaging protocols, patient populations, and anatomical characteristics was used for training and evaluation. Performance metrics, robustness to variations in image quality, and age-related associations with PVS burden were rigorously evaluated against established RORPO-based methods.
RESULT: SegcsvdPVS achieved high sensitivity for basal ganglia PVS (SNS = 0.81 ± 0.13) and identified significantly larger volumes (86.1 ± 67.1 mm3) compared to human tracers (47.2 ± 26.5 mm3, 48.6 ± 28.4 mm3. Our tool demonstrated strong age-related correlations with PVS volumes across three diverse datasets (TEST: r = 0.41, CI = [0.03, 0.68]; ADNI: r = 0.38, CI = [0.30, 0.46]; CAHHM: r = 0.41, CI = [0.35, 0.46]). Although similar but weaker trends were observed for non-basal ganglia PVS, segcsvdPVS demonstrated superior robustness to variations in contrast and noise across both regions, with minimal changes in age-related correlations (∆r ≤ 0.08) compared to the RORPO-based methods (∆r ≤ 0.39).
CONCLUSION: SegcsvdPVS is a reliable tool for PVS segmentation, particularly in basal ganglia regions, offering superior sensitivity, robustness to imaging variability, and enhanced detection of biologically relevant age-related associations. These findings support its application in large-scale studies and clinical research to advance our understanding of PVS contributions to CSVD and dementia.
PMID:41453363 | DOI:10.1002/alz70856_104470