Reliability-based robust multi-atlas label fusion for brain MRI segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Neuroimaging

Background:

• Multi-atlas based segmentation is crucial for structural magnetic resonance (MR) image analysis.
• Existing label fusion methods often overlook voxel reliability differences and the potential for high-reliability voxels to refine low-reliability ones.

Purpose of the Study:

• To propose a general reliability-based robust label fusion framework for multi-atlas MR image segmentation.
• To enhance segmentation accuracy by incorporating label and spatial reliability for refining low-reliability voxels.

Main Methods:

• An initial segmentation is performed using conventional multi-atlas label fusion.
• Two types of reliability (label and spatial) are defined and estimated for each voxel based on initial segmentation.
• A refinement step utilizes high-reliability voxels to improve the segmentation of low-reliability voxels.

Main Results:

• The proposed framework was integrated into four established methods (LWV, PBM, JLF, SPBM), creating ls-LWV, ls-PBM, ls-JLF, and ls-SPBM.
• Validation on brain MR images from NIREP, LONI-LPBA40, and ADNI datasets showed superior performance.
• The label-spatial reliability-based methods outperformed state-of-the-art techniques in multi-atlas image segmentation.

Conclusions:

• The developed label-spatial reliability-based framework offers a robust approach to multi-atlas MR image segmentation.
• This method effectively addresses limitations of existing techniques by incorporating voxel reliability for improved accuracy.
• The enhanced fusion methods demonstrate significant improvements in segmenting regions of interest in brain MR images.