Objectives To develop and validate a clinically applicable deep learning framework for automated segmentation of intracranial and carotid vessel walls in black-blood magnetic resonance vessel wall imaging (MR-VWI). Materials and methods In this retrospective multi-center study, 193 patients (mean age: 60.2 ± 4.3 years) from five hospitals underwent high-resolution black-blood MR-VWI. A deep learning segmentation framework was developed incorporating three key innovations: (1) polar coordinate mapping, (2) a feature-sharing padding strategy, and (3) a polar Dice loss function. Manual expert annotations served as the reference standard for training and evaluation. Model performance was assessed using Dice similarity coefficients (DSC), Hausdorff distances (HD), and area differences (AD) for both lumen and vessel wall regions. External validation was performed on an independent multi-center test set from four external institutions, and the publicly available MICCAI 2021 Vessel Wall Segmentation Challenge dataset. Gradient-weighted Class Activation Mapping (Grad-CAM) was used for interpretability. Results On the external test set, the model achieved DSCs of 0.928 (outer wall area), 0.936 (lumen area), and 0.844 (vessel wall region). It significantly outperformed four benchmark networks in boundary and area accuracy (all p < 0.05). On the public MICCAI dataset, it achieved the highest vessel wall DSC (0.782) and the lowest lumen and wall area errors. Grad-CAM confirmed that the model consistently focused on anatomically relevant vessel wall boundaries. Conclusion This deep learning-based method enables accurate and reproducible vessel wall segmentation in clinical black-blood MR-VWI, offering a practical solution to streamline cerebrovascular risk assessment and support decision-making in stroke prevention and monitoring. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Can a clinically oriented deep learning framework provide accurate, generalizable vessel wall segmentation across diverse vascular territories in black-blood MR imaging? <Emphasis Type="BoldItalic">Findings</Emphasis> The proposed model achieved superior segmentation performance compared to four benchmark networks, with consistent accuracy on multi-center and public test datasets. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> Reliable vessel wall segmentation may support more objective quantification of intracranial atherosclerosis, enabling early diagnosis, treatment planning, and longitudinal monitoring of high-risk stroke patients. Graphical Abstract

Clinically oriented deep learning framework for automated vessel wall segmentation in black-blood MRI: a multi-center study

期刊论文
发表日期：

2025年11月22日

查看全文

Xuetong Tao,
Shuai Shen,
Long Yang,
Kexin Chen,
Hui Yang,
Greta S. P. Mok,
Lin Jia,
Xin Liu,
Dong Liang,
Zhanli Hu,
Hairong Zheng,
Na Zhang

摘要

Objectives

To develop and validate a clinically applicable deep learning framework for automated segmentation of intracranial and carotid vessel walls in black-blood magnetic resonance vessel wall imaging (MR-VWI).

Materials and methods

In this retrospective multi-center study, 193 patients (mean age: 60.2 ± 4.3 years) from five hospitals underwent high-resolution black-blood MR-VWI. A deep learning segmentation framework was developed incorporating three key innovations: (1) polar coordinate mapping, (2) a feature-sharing padding strategy, and (3) a polar Dice loss function. Manual expert annotations served as the reference standard for training and evaluation. Model performance was assessed using Dice similarity coefficients (DSC), Hausdorff distances (HD), and area differences (AD) for both lumen and vessel wall regions. External validation was performed on an independent multi-center test set from four external institutions, and the publicly available MICCAI 2021 Vessel Wall Segmentation Challenge dataset. Gradient-weighted Class Activation Mapping (Grad-CAM) was used for interpretability.

Results

On the external test set, the model achieved DSCs of 0.928 (outer wall area), 0.936 (lumen area), and 0.844 (vessel wall region). It significantly outperformed four benchmark networks in boundary and area accuracy (all p < 0.05). On the public MICCAI dataset, it achieved the highest vessel wall DSC (0.782) and the lowest lumen and wall area errors. Grad-CAM confirmed that the model consistently focused on anatomically relevant vessel wall boundaries.

Conclusion

This deep learning-based method enables accurate and reproducible vessel wall segmentation in clinical black-blood MR-VWI, offering a practical solution to streamline cerebrovascular risk assessment and support decision-making in stroke prevention and monitoring.

Key Points

Question Can a clinically oriented deep learning framework provide accurate, generalizable vessel wall segmentation across diverse vascular territories in black-blood MR imaging?

Findings The proposed model achieved superior segmentation performance compared to four benchmark networks, with consistent accuracy on multi-center and public test datasets.

Clinical relevance Reliable vessel wall segmentation may support more objective quantification of intracranial atherosclerosis, enabling early diagnosis, treatment planning, and longitudinal monitoring of high-risk stroke patients.

Graphical Abstract