Multi-stage deep learning architecture for carotid artery segmentation and stenosis evaluation: comparative study...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• High-resolution magnetic resonance imaging (HR-MRI) is a non-invasive method for evaluating carotid atherosclerosis stenosis.
• Manual segmentation and stenosis evaluation are time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate a novel automated architecture for segmenting extracranial carotid arteries and evaluating stenosis using HR-MRI.
• To compare the proposed architecture's performance against manual segmentation and digital subtraction angiography (DSA).

Main Methods:

• A dataset of 641 stenotic arteries from 422 patients was used, divided into training-validation and independent test sets.
• An external validation set from a fourth hospital was also included.
• The architecture's performance was evaluated using Dice similarity coefficients and stenosis evaluation accuracy.

Main Results:

• The automated architecture achieved high consistency with manual segmentation (Dice scores of 0.97 ± 0.01 and 0.96 ± 0.01).
• Stenosis evaluation accuracy reached 0.88 on the independent test set and 0.86 on the external validation set.
• Performance was comparable to DSA diagnostic criteria.

Conclusions:

• The proposed architecture provides accurate and reliable automated segmentation and stenosis evaluation of carotid arteries from HR-MRI.
• It significantly reduces diagnostic time and inter-observer variability compared to manual methods.
• This intelligent tool shows promise for diagnosing head and neck atherosclerotic disease and assessing stroke risk.