A Region and Category Confidence-Based Multi-Task Network for Carotid Ultrasound Image Segmentation and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• Carotid plaque segmentation and classification are crucial for atherosclerosis management and stroke risk assessment.
• Existing deep learning methods often increase complexity or overlook task interdependencies, limiting performance.

Purpose of the Study:

• To propose a multi-task learning framework (RCCM-Net) for enhanced carotid plaque segmentation and classification in ultrasound images.
• To leverage the correlation between segmentation and classification tasks for improved performance.

Main Methods:

• Developed RCCM-Net, a multi-task learning framework incorporating a region confidence module (RCM) and a sample category confidence module (CCM).
• Trained and evaluated the model on 1270 2D ultrasound carotid plaque images.
• Compared RCCM-Net against various single-task and multi-task deep learning networks.

Main Results:

• RCCM-Net achieved 85.82% classification accuracy and 84.92% Dice-similarity-coefficient for segmentation.
• The proposed method outperformed existing single-task and multi-task networks.
• Ablation studies confirmed the significant contribution of both RCM and CCM modules.

Conclusions:

• RCCM-Net effectively exploits the relationship between carotid plaque segmentation and classification tasks.
• The framework offers improved performance for clinical applications in carotid plaque analysis.
• This approach holds promise for enhancing clinical trials and practice related to atherosclerosis.