Neuroradiology · 2026
A comprehensive framework for automated segmentation of perivascular spaces in brain MRI with the nnU-Net.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Pham William, Jarema Alexander, Rim Donggyu, Chen Zhibin, Khlif Mohamed, Macefield Vaughan, Henderson Luke, Brodtmann Amy
Аннотация
BACKGROUND: Enlargement of perivascular spaces (PVS) is common in cerebral small vessel disease, Alzheimer’s disease, and Parkinson’s disease, reflecting impaired clearance pathways. While MRI provides a means to quantify perivascular spaces, manual annotation remains time-consuming and labour-intensive. Thus, there is a need for accurate automated MRI-based PVS segmentation methods.
AIM: To optimise the nnU-Net, a deep-learning framework, for PVS segmentation.
METHODS: 30 T1-weighted (T1w) MRI images acquired on three different scanners were used. PVS in the white matter (WM) and basal ganglia (BG) were manually labelled via a sparse annotation strategy and used to optimise the nnU-Net for T1w PVS segmentation. The same pipeline was applied to T2-weighted (T2w) images. Additionally, we trained T1w+FLAIR and T2w+FLAIR models to simultaneously segment PVS and white matter hyperintensities (WMH). Performance was assessed with 5-fold cross validation using the Dice similarity coefficient (DSC).
RESULTS: A voxel-spacing agnostic model (mean DSC = 64.3 ± 3.3%) outperformed models that resampled images to a common resolution (DSC = 40.5–55%). Training on PVS segmentations derived from preprocessed T1w images substantially improved performance (DSC = 78.3 ± 1.7%). The T2w model performed best overall (DSC = 84.7 ± 1.3%), especially for WM-PVS (DSC = 90.4 ± 0.9%) compared to BG-PVS (DSC = 79.1 ± 2%). Multimodal models achieved DSCs of 75.6 ± 3.4% (T1w+FLAIR) and 77.5 ± 2.7% (T2w+FLAIR).
CONCLUSIONS: Our deep learning models provide a robust framework for automated PVS quantification across MRI modalities.
SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00234-026-03993-y.
AIM: To optimise the nnU-Net, a deep-learning framework, for PVS segmentation.
METHODS: 30 T1-weighted (T1w) MRI images acquired on three different scanners were used. PVS in the white matter (WM) and basal ganglia (BG) were manually labelled via a sparse annotation strategy and used to optimise the nnU-Net for T1w PVS segmentation. The same pipeline was applied to T2-weighted (T2w) images. Additionally, we trained T1w+FLAIR and T2w+FLAIR models to simultaneously segment PVS and white matter hyperintensities (WMH). Performance was assessed with 5-fold cross validation using the Dice similarity coefficient (DSC).
RESULTS: A voxel-spacing agnostic model (mean DSC = 64.3 ± 3.3%) outperformed models that resampled images to a common resolution (DSC = 40.5–55%). Training on PVS segmentations derived from preprocessed T1w images substantially improved performance (DSC = 78.3 ± 1.7%). The T2w model performed best overall (DSC = 84.7 ± 1.3%), especially for WM-PVS (DSC = 90.4 ± 0.9%) compared to BG-PVS (DSC = 79.1 ± 2%). Multimodal models achieved DSCs of 75.6 ± 3.4% (T1w+FLAIR) and 77.5 ± 2.7% (T2w+FLAIR).
CONCLUSIONS: Our deep learning models provide a robust framework for automated PVS quantification across MRI modalities.
SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00234-026-03993-y.