Optimizing boundary detection via Simulated Search with applications to multi-modal heart segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Biomedical engineering

Background:

• Accurate medical image segmentation relies on robust boundary detection.
• Current methods assess boundary detection globally, limiting local optimization.
• Anatomical variability poses challenges for automated segmentation initialization.

Purpose of the Study:

• To introduce a novel method for local assessment of boundary detection in medical image segmentation.
• To enable landmark-specific optimization of boundary detection during model training.
• To improve the accuracy and automation of medical image segmentation.

Main Methods:

• Proposed the Simulated Search method for local boundary detection assessment.
• Evaluated performance by estimating geometric boundary detection error for individual model landmarks.
• Demonstrated optimization of boundary detection per landmark during model training.

Main Results:

• Simulated Search improved capture range and accuracy compared to traditional training.
• Successfully applied to cardiac image segmentation across multiple modalities (CT, MRI, 3D X-ray angiography).
• Achieved average segmentation errors of 0.8 mm for cardiac chambers/vessels and 1.3 mm for left atrium/pulmonary veins.

Conclusions:

• The Simulated Search method enables effective local optimization of boundary detection.
• This approach facilitates fully automatic and accurate segmentation in diverse medical imaging scenarios.
• The method aids in identifying suitable features for new segmentation tasks and supports multi-modal segmentation.