CMIS: A Class-Aware Multi-Structure Instance Segmentation Model for Fetal Brain Ultrasound Images With Fuzzy...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Fetal Medicine

Background:

• Fetal anatomical segmentation in ultrasound is crucial for diagnosis and measurement.
• Current methods are limited to specific planes or structures and struggle with fuzzy regions.
• Obstetricians require multi-plane, multi-structure analysis for comprehensive diagnosis.

Purpose of the Study:

• To introduce a real-time segmentation method, Class-aware Multi-structure Instance Segmentation (CMIS), for 19 key fetal brain structures across 3 planes.
• To enhance fetal brain-disease diagnosis by addressing limitations of existing segmentation techniques.
• To improve segmentation accuracy in challenging cases with fuzzy boundaries and varying scales.

Main Methods:

• Developed CMIS utilizing instance information and class-aware attention for computational efficiency and detailed insights.
• Implemented cross-layer and multi-scale fusion to generate detailed prototypes.
• Introduced a fuzzy region-based constraint loss and random box perturbation during training to enhance robustness.

Main Results:

• CMIS achieved a mean Dice score of 83.41% at 37 FPS on a fetal brain dataset, outperforming 13 baselines.
• The method demonstrated strong performance on a fetal heart ultrasound dataset with a mean Dice score of 85.73%.
• CMIS effectively segments complex anatomical structures in ultrasound, showing potential for real-time clinical applications.

Conclusions:

• CMIS offers a robust and efficient solution for segmenting multiple fetal brain structures in ultrasound images.
• The method's ability to handle fuzzy regions and its real-time performance make it suitable for clinical applications.
• Further investigation is needed for generalization to abnormal cases and diverse datasets beyond 2D normal standard planes.