Automated Screening Network for Fetal Closed Spina Bifida With Semantic Enhancement and Projected Attention

Area of Science:

• Medical imaging
• Artificial intelligence
• Developmental biology

Background:

• Closed spina bifida is a rare fetal disease with subtle ultrasound signs, often leading to misdiagnosis.
• Accurate diagnosis relies heavily on sonographer expertise, highlighting the need for automated tools.

Purpose of the Study:

• To develop a novel semantic enhancement framework with projected attention for automated closed spina bifida screening.
• To improve precise landmark detection in fetal ultrasound images for early disease identification.

Main Methods:

• Utilized a multi-granularity deep supervision and voting mechanism for point-specific feature generation and saliency map reconstruction.
• Employed a coordinate attention projection module to convert 2D probability maps to 1D vectors for precise coordinate regression with low complexity.
• Developed an intelligent system for automated fetal spine counting and anatomical measurement.

Main Results:

• The proposed method effectively reduces interference from image noise while preserving semantic information.
• Achieved precise coordinate regression with low computational complexity.
• Demonstrated clear advantages in computational complexity and accuracy compared to advanced baselines on multiple datasets.

Conclusions:

• The developed framework offers significant potential for clinical application in automated fetal spine analysis.
• Facilitates early disease warnings through accurate anomaly identification.
• Represents a significant advancement in AI-assisted diagnostic tools for fetal developmental disorders.