Dual-task ultrasound spine transverse vertebrae segmentation network with contour regularization.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Radiology

Background:

• 3D ultrasound is a common, radiation-free method for scoliosis diagnosis.
• Accurate spine curvature measurement requires precise segmentation of vertebrae from 2D coronal images.
• Current segmentation methods face challenges in boundary definition and interference.

Purpose of the Study:

• To develop an advanced deep learning model for automatic segmentation of transverse vertebrae in 3D ultrasound images.
• To improve the accuracy and robustness of vertebrae segmentation for precise scoliosis assessment.
• To enhance the measurement of spine curvature angles by refining vertebral feature identification.

Main Methods:

• Proposed a dual-task ultrasound transverse vertebrae segmentation network (D-TVNet) based on the U-Net architecture.
• Integrated an auxiliary shape regularization network to improve contour segmentation and anti-interference capabilities.
• Incorporated an atrous spatial pyramid pooling (ASPP) module and fused auxiliary network features to enhance feature extraction and boundary segmentation.

Main Results:

• The D-TVNet achieved a best Dice score of 86.68% and a mean Dice score of 86.17% via cross-validation.
• Demonstrated a significant improvement of 5.17% over the baseline U-Net model.
• Validated the effectiveness of the dual-task approach and feature fusion strategies for vertebrae segmentation.

Conclusions:

• The D-TVNet provides a promising automatic solution for ultrasound spine bone segmentation.
• The proposed method enhances the precision of vertebrae segmentation, crucial for accurate scoliosis diagnosis.
• This advancement contributes to the development of more effective and reliable 3D ultrasound-based diagnostic tools for spinal conditions.