BECM-Net: A Multi-granularity Collaborative Framework for Semi-Supervised Fetal Ultrasound Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate segmentation of fetal ultrasound (US) images is crucial for clinical assessments like measuring the Angle of Progression (AoP) and fetal head descent.
• Conventional semi-supervised learning (SSL) methods struggle with blurred boundaries and limited consistency enforcement in US image segmentation.

Purpose of the Study:

• To develop an advanced deep learning framework, BECM-Net, for improved fetal US image segmentation.
• To address challenges in pseudo-labeling accuracy and consistency in semi-supervised learning for ultrasound segmentation.

Main Methods:

• Proposed the Boundary-Enhanced Collaborative Multi-granularity Network (BECM-Net), a unified framework optimizing pixel, region, and structure-level representations.
• Introduced DirDiff-Conv for enhanced boundary perception and texture representation at the pixel level.
• Implemented Uncertainty-Confidence Aligned Mix (UCA-Mix) for uncertainty-guided region-level mixing and ContourRefine for structure-level contour modeling.

Main Results:

• BECM-Net achieved state-of-the-art performance on fetal ultrasound datasets.
• Demonstrated significant improvements in segmenting challenging regions with ambiguous pubic symphysis and fetal head boundaries.
• Showcased more reliable supervision and robust feature learning with limited annotations.

Conclusions:

• BECM-Net offers a robust solution for fetal ultrasound segmentation by integrating multi-granularity modeling.
• The proposed methods effectively enhance boundary perception, reduce pseudo-label noise, and enforce structural consistency.
• BECM-Net holds promise for improving clinical assessments of labor progression and fetal well-being.