Deep learning-based fully automatic screening of carotid artery plaques in computed tomography angiography: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• Carotid artery plaques (CAPs) are crucial indicators of cardiovascular risk.
• Accurate detection and classification of CAPs on computed tomography angiography (CTA) are essential for patient management.
• Current manual analysis can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop and validate a deep learning (DL) algorithm for automated detection and classification of CAPs.
• To assess the DL model's performance using various metrics on independent validation datasets.

Main Methods:

• A retrospective study involving 400 patients with CTA scans.
• Development of a two-step DL system using modified 3D-UNet for segmentation and ResUNet for detection/classification.
• Validation using internal (300 patients) and external (100 patients) datasets, with radiologist consensus as ground truth.

Main Results:

• The DL model achieved 83.4% sensitivity (internal) and 78.9% (external) for CAP detection.
• Good overall accuracy indicated by F1-scores (0.764 internal, 0.769 external) and area under fROC curves (0.756 internal, 0.738 external).
• Effective ternary classification of CAPs (noncalcified, mixed, calcified) with Cohen's kappa of 0.728 (internal) and 0.703 (external).

Conclusions:

• A fully automated DL algorithm for CAP detection and ternary classification is feasible.
• The developed DL system demonstrates robust performance on both internal and external validation datasets.
• This automated approach has the potential to assist radiologists and improve workflow efficiency.