TL;DR
This paper introduces a novel multi-stage deep learning architecture for automatically segmenting carotid arteries and evaluating stenosis from high-resolution MRI (HR-MRI). This approach achieves accuracy comparable to manual segmentation and high consistency with Digital Subtraction Angiography (DSA), offering a faster, more reliable, and less variable tool for diagnosing atherosclerotic disease and assessing stroke risk.
What Problem Does This Paper Address?
Cerebrovascular diseases, particularly ischemic strokes, are a major health concern, with carotid artery atherosclerosis being a leading cause. Accurately identifying and quantifying stenosis in the carotid arteries is crucial for predicting stroke risk and guiding treatment. While Digital Subtraction Angiography (DSA) is the gold standard, it's invasive and uses radiation. High-resolution Magnetic Resonance Imaging (HR-MRI) offers a non-invasive alternative, but manual analysis of HR-MRI images for segmentation and stenosis assessment is time-consuming, labor-intensive, and prone to inter-observer variability. Existing automated methods often struggle with the complex anatomy of carotid arteries, leading to inconsistent segmentation. This paper aims to bridge this gap by developing an automated, accurate, and efficient deep learning system for carotid artery segmentation and stenosis evaluation using HR-MRI.
Methods
The authors propose a three-stage deep learning architecture:
- Carotid Artery Localization Module (LC): This module uses a modified 2D U-Net to identify a Region of Interest (ROI) containing the extracranial carotid artery from 2D axial HR-MRI slices. This step narrows the focus for subsequent analysis, improving efficiency and accuracy.
- Artery Automatic Segmentation Module: This stage employs a modified 3D U-Net architecture to perform precise 3D segmentation of the carotid artery within the identified ROI. A residual connection is incorporated into the encoder to enhance training efficiency and segmentation accuracy, addressing issues like gradient vanishing.
- Stenosis Automatic Diagnosis Module: This module first straightens the segmented artery to reduce computational complexity and then quantifies stenosis. It identifies the reference vessel diameter (RVD) and minimum lumen diameter (MLD) to calculate the stenosis degree based on NASCET criteria.
The system incorporates image preprocessing (resampling, Z-Score normalization) and data augmentation techniques to improve model robustness and generalization.
Datasets
The study utilized a retrospective, multicenter dataset:
- Training-validation set: 372 patients with 545 lesions.
- Independent test set: 50 patients with 96 lesions.
- External validation set: 89 patients with 168 lesions, collected prospectively from a fourth hospital.
All HR-MRI images were converted from DICOM to NIfTI format. DSA was used as the reference standard for validating stenosis evaluation.
Key Insights and Ideas
- Multi-stage approach is key: Breaking down the complex task into localization, segmentation, and stenosis evaluation allows for more focused and accurate processing at each step.
- Localization as a precursor: The initial 2D localization module significantly improves the efficiency and accuracy of the subsequent 3D segmentation by defining a precise ROI.
- Residual connections for robustness: Incorporating residual blocks in the 3D U-Net encoder helps mitigate gradient vanishing and improves the stability and performance of the segmentation model.
- Straightening for accurate stenosis measurement: Straightening the segmented artery is a crucial step to overcome the challenges posed by curved anatomy, enabling more reliable quantitative stenosis assessment.
- High consistency with gold standard: The architecture demonstrates strong agreement with manual segmentation and DSA diagnostic criteria, suggesting its clinical viability.
Key Results
- Segmentation Performance: The proposed architecture achieved high segmentation accuracy, with mean Dice similarity coefficients (DSC) of 0.97 ± 0.02 on the training-validation set, 0.97 ± 0.02 on the independent test set, and 0.95 ± 0.05 on the external validation set. Concordance correlation coefficients (CCC) between ground truth and segmentation volumes were high (e.g., 0.985 for training-validation, 0.979 for independent test, 0.963 for external validation).
- Stenosis Evaluation Performance: Stenosis evaluation accuracies were 0.88 on the independent test set and 0.86 on the external validation set.
- Comparison with SOTA: The proposed method outperformed existing state-of-the-art models in segmentation tasks, achieving higher DSC and IOU while maintaining competitive ASSD and HD95 values.
- Ablation Study: Removing the localization module (LC) significantly degraded performance, highlighting its critical contribution.
Limitations and Future Work
- Generalizability: The authors acknowledge the need for further validation on more diverse datasets to ensure broader applicability, especially for less common causes of carotid stenosis (e.g., arteritis, dissection).
- Missed Segmentation: A limitation was observed in one case where an artery in the most stenotic region was not fully captured, indicating potential for improvement.
- Computational Resources: The sequential two-phase architecture requires significant computational resources, and future work will focus on optimizing inference speed.
- Mild/Moderate Stenosis Accuracy: The diagnostic module showed limitations in diagnosing mild to moderate stenosis, with higher error rates compared to severe stenosis.
- ROI Definition: The coarse localization module sometimes included adjacent structures (like the jugular vein), potentially affecting lumen and reference vessel diameter measurements. Future work will refine this module to better separate carotid arteries from surrounding tissues.
- Expanded Analysis: Future research aims to incorporate vessel wall segmentation and detailed plaque burden/vulnerability assessments to leverage HR-MRI's full capabilities.
Summary updated