Deep Learning for Carotid Plaque Segmentation using a Dilated U-Net Architecture.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Ultrasound

Background:

• Carotid plaque segmentation is crucial for assessing stenosis severity.
• Deep learning models offer potential for automated image analysis.
• Accurate segmentation aids in cardiovascular risk stratification.

Purpose of the Study:

• To evaluate deep learning models for carotid plaque segmentation in ultrasound B-mode images.
• To compare standard U-Net and dilated U-Net architectures.
• To assess the impact of bounding box guidance on segmentation performance.

Main Methods:

• Utilized a dataset of 101 severely stenotic carotid plaque patients.
• Implemented and compared fully automatic and semi-automatic U-Net and dilated U-Net models.
• Quantified performance degradation due to bounding box errors using Dice coefficients.

Main Results:

• Semi-automatic segmentation with bounding boxes significantly improved performance (Dice: 0.83-0.84) over automatic methods (Dice: 0.48-0.55).
• Dilated U-Net showed slightly better automatic segmentation (Dice: 0.55) than standard U-Net (Dice: 0.48).
• A 5% error in bounding box dimensions reduced Dice coefficients to 0.79-0.80.

Conclusions:

• Bounding box-guided semi-automatic deep learning segmentation enhances carotid plaque analysis accuracy.
• Deep learning models show promise for automated carotid plaque segmentation, with potential for further improvement.
• Robustness to bounding box inaccuracies is critical for clinical translation.