A dual-modal dynamic contour-based method for cervical vascular ultrasound image instance segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Vascular Ultrasound

Background:

• Accurate segmentation of carotid artery and jugular vein is crucial for vascular health assessment.
• Current single-modal deep learning methods face limitations in capturing complex vascular structures.

Purpose of the Study:

• To develop a novel dual-modal dynamic contour-based instance segmentation method.
• To integrate ultrasound and optical flow images for enhanced segmentation accuracy.
• To compare the proposed method against single-modal deep learning networks.

Main Methods:

• Collected and curated a dataset of 2432 carotid artery and jugular vein ultrasound images.
• Generated optical flow images from ultrasound data to highlight vessel contours.
• Developed a dual-stream information fusion module and a learnable contour initialization technique.

Main Results:

• Achieved a mean average precision of 0.814 for bounding box detection and 0.842 for mask segmentation.
• Demonstrated smoother segmentation boundaries for blood vessels in qualitative analysis.
• Validated performance on a self-built dataset of vascular ultrasound images.

Conclusions:

• The dual-modal network effectively leverages complementary features from ultrasound and optical flow images.
• The proposed method offers more accurate segmentation of carotid artery and jugular vein compared to single-modal approaches.
• This technique shows significant potential for precise and reliable medical image analysis in vascular applications.