Label image constrained multiatlas selection.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Multi-atlas based methods are standard for medical image segmentation.
• Atlas selection and combination are critical for performance in these methods.
• Manifold learning shows promise for atlas selection but faces challenges with complex structures.

Purpose of the Study:

• To propose a label image constrained atlas selection method for improved medical image segmentation.
• To address the difficulty of accurate atlas selection using raw image distances on manifolds.
• To develop a novel atlas combination weighting method based on manifold subspace analysis.

Main Methods:

• Proposed a label image constrained manifold projection for raw images.
• Utilized label images to guide the manifold projection of T2w MRI data.
• Developed a novel atlas combination weighting strategy by analyzing data point distribution in the manifold subspace.

Main Results:

• The proposed method demonstrated improved atlas selection accuracy for prostate structures.
• Experimental results showed selected atlases were closer to the target structure.
• Achieved more accurate prostate segmentation compared to existing methods on T2w MRI.

Conclusions:

• Label image constraints effectively improve manifold-based atlas selection in medical imaging.
• The novel weighting method enhances atlas combination for better segmentation outcomes.
• The proposed approach offers a significant advancement for T2w MRI prostate segmentation.