Deep Learning-Based Carotid Plaque Segmentation from B-Mode Ultrasound Images

Area of Science:

• Medical Imaging
• Cardiovascular Disease Research
• Artificial Intelligence in Medicine

Background:

• Carotid ultrasound measures total plaque area (TPA) to assess carotid atherosclerosis burden and treatment response.
• Manual plaque segmentation for TPA calculation is time-consuming and requires extensive observer training.

Purpose of the Study:

• To develop an automated plaque segmentation method for generating TPA measurements from longitudinal carotid ultrasound images.
• To utilize a modified U-Net deep learning model for accurate and efficient TPA quantification.

Main Methods:

• A modified U-Net deep learning model was employed for automated plaque segmentation.
• The dataset comprised 510 plaques from 144 patients, with Monte Carlo cross-validation for training and testing.
• Two U-Net models were trained using manual delineations from two observers as ground truth.

Main Results:

• The automated segmentation showed strong agreement with manual measurements (Pearson's r = 0.989 and 0.987, p < 0.0001).
• Mean TPA differences between automated and manual methods were small (0.05 ± 7.13 mm² and 0.8 ± 8.7 mm²).
• Segmentation time was significantly reduced to an average of 8.3 ± 3.1 ms per plaque.

Conclusions:

• The deep learning-based method provides accurate and rapid TPA measurements for carotid atherosclerosis.
• The high agreement with manual segmentations supports its use in monitoring plaque progression and regression.
• This automated approach offers a time-efficient alternative to manual segmentation in clinical practice.