Ultrasound segmentation analysis via distinct and completed anatomical borders

Area of Science:

• Medical Imaging Analysis
• Deep Learning in Healthcare
• Ultrasound Segmentation

Background:

• Accurate segmentation of ultrasound images is crucial for medical diagnosis and treatment.
• Current deep learning models lack transparency in how they define boundaries.
• Ultrasound imaging presents unique challenges for border delimitation.

Purpose of the Study:

• To analyze ultrasound segmentation networks by examining their learned borders.
• To differentiate between distinct and completed borders in ultrasound images.
• To understand network attention mechanisms in border definition.

Main Methods:

• Splitting ultrasound image borders into distinct and completed categories.
• Utilizing Grad-CAM to visualize network attention on split borders.
• Quantifying prediction accuracy for distinct and completed borders.
• Experimenting on diverse ultrasound datasets (leg, thyroid, nerves, prostate).

Main Results:

• Networks achieved approximately 10% better performance on completed borders than distinct ones.
• Network performance mirrors clinical observations, struggling with less visible areas.
• Seg-Grad-CAM analysis revealed distinct border focus on shiny structures, while completed borders utilize landmarks.
• Attention mechanisms influenced network performance variations.

Conclusions:

• Studying ultrasound borders requires distinct approaches compared to MRI or CT.
• Differentiating between distinct and completed borders reveals network learning quality.
• A 3D leg ultrasound dataset is publicly released to aid research.