Rapid vessel segmentation and reconstruction of head and neck angiograms from MR vessel wall images

Jin Zhang, Wen Wang, Jinhua Dong, Xiong Yang, Shuwei Bai, Jiaqi Tian, Bo Li, Xiao Li, Jianjian Zhang, Hangyu Wu, Xiaoxi Zeng, Yongqiang Ye, Shenghao Ding, Jieqing Wan, Ke Wu, Yufei Mao, Cheng Li, Na Zhang, Jianrong Xu, Yongming Dai, Feng Shi, Beibei Sun, Yan Zhou, Huilin ZhaoNpj Digit. Medicine 2025 · DOI

TL;DR

This paper introduces "VWI Assistant," a deep learning platform that automates the segmentation and reconstruction of head and neck angiograms from MR vessel wall images. It significantly reduces processing time and labor costs, achieving high accuracy comparable to manual methods, and has demonstrated rapid clinical adoption.

What Problem Does This Paper Address?

Three-dimensional magnetic resonance vessel wall imaging (3D MR-VWI) is a powerful tool for understanding cerebrovascular diseases like stroke. However, its clinical use is hampered by the extremely labor-intensive and time-consuming manual post-processing required for image segmentation and reconstruction. This bottleneck limits the scalability of 3D MR-VWI for routine clinical diagnosis and large-scale research, potentially delaying diagnoses and increasing healthcare costs. The paper aims to overcome this by developing an automated, efficient, and accurate system for this crucial step.

Methods

The core of the solution is "VWI Assistant," a deep learning platform built on convolutional neural networks (CNNs). It employs a two-phase approach:

  • Phase I: Vessel Segmentation and Recognition: This phase uses enhanced multi-channel and single-channel CNNs to automatically segment both black-blood and bright-blood MR sequences. The segmentation results from different sequences are fused to generate a final, accurate segmentation and recognition of vessel segments.
  • Phase II: Fast Curved Planar Reformation (CPR): This phase extracts principal arterial centerlines by skeletonizing the segmented vessels. A distance field-driven path tracing algorithm then connects and extends these centerlines to form complete vascular pathways. Lumen segmentation data is integrated for precise centerline correction, especially in black-blood sequences.

The framework is designed to be flexible, handling incomplete imaging sequences by automatically adapting to available data. Data augmentation techniques (displacement, truncation, left-right reversal, simulation of occluded vessels) were used to increase dataset size and improve model robustness.

Datasets

The study utilized a large, multi-center dataset comprising 1981 patients undergoing 3D MR-VWI. This dataset was split into:

  • Training/Validation/Testing: 748 scans were used for model development.
  • Clinical Evaluation: 134 patients were used to assess the system's performance against manual reconstruction.
  • Clinical Application: 1099 patients were used to evaluate real-world adoption and efficiency.

The data came from four institutions and included scans from various manufacturers (Philips, Siemens, United Imaging, GE) and diverse clinical scenarios (normal arteries, atherosclerotic stenosis, occlusion, aneurysm).

Key Insights and Ideas

  • Deep Learning for Automation: The paper demonstrates that deep learning can effectively automate a complex and time-consuming medical imaging task, significantly improving efficiency.
  • Multi-Sequence Integration: Combining information from different MR sequences (T1WI, TOF-MRA, T1WI-CE) is crucial for achieving the best segmentation and reconstruction performance, leveraging the unique strengths of each sequence.
  • Real-World Clinical Impact: The rapid adoption rate (reaching 100% utilization within 12 months) and consistent qualification rates in a clinical setting highlight the practical value and user acceptance of the VWI Assistant.
  • Standardization of Diagnostics: AI-driven processing offers greater consistency and objectivity compared to manual methods, reducing inter- and intra-reader variability in measurements like stenosis rate.

Key Results

  • Accuracy: Achieved a 92.9% qualified rate for reconstructed images, comparable to manual methods (93.5%).
  • Speed: Reduced processing time by over 90%, from an average of 10-12 minutes per case for manual methods to approximately 1 minute for VWI Assistant.
  • Efficiency: Significantly reduced the number of clicks required for processing (from ~400 to ~16).
  • Clinical Adoption: Real-world deployment showed a rapid increase in utilization, reaching 100% within 12 months.
  • Inter-/Intra-Reader Agreement: VWI Assistant postprocessing showed excellent agreement (ICC > 0.900) for lesion measurements, even better than manual methods.

Limitations and Future Work

  • Distal Arteries: The model struggles with segmenting and reconstructing small, distal intracranial arteries (e.g., branches post M2 segment of the middle cerebral artery) due to its primary design for larger vessels.
  • Complex Pathologies: Accuracy can be reduced in complex vascular pathologies like fibromuscular dysplasia or Moyamoya disease, sometimes requiring manual correction.
  • Rare Variants/Low-Quality Data: A small percentage of cases (over 5%) still require manual adjustments due to insufficient representation of rare anatomical variants or low-quality training data.
  • Future Work: The authors suggest that continued learning and improvement of the AI model will further decrease the need for manual correction. They also highlight the potential for integrating this technology with AI tools for lesion detection and analysis.

Summary updated

BibTeX
@article{zhang2025rapid,
  author = {Jin Zhang and Wen Wang and Jinhua Dong and Xiong Yang and Shuwei Bai and Jiaqi Tian and Bo Li and Xiao Li and Jianjian Zhang and Hangyu Wu and Xiaoxi Zeng and Yongqiang Ye and Shenghao Ding and Jieqing Wan and Ke Wu and Yufei Mao and Cheng Li and Na Zhang and Jianrong Xu and Yongming Dai and Feng Shi and Beibei Sun and Yan Zhou and Huilin Zhao},
  title = {{Rapid vessel segmentation and reconstruction of head and neck angiograms from MR vessel wall images}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-01866-x},
}
Figure