A GENERATIVE MODEL FOR MULTI-ATLAS SEGMENTATION ACROSS MODALITIES.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning for medical imaging

Background:

• Current multi-atlas segmentation methods rely on voxel intensity similarity, limiting cross-modality applications.
• Inconsistencies in voxel intensity across different MRI datasets hinder traditional label fusion techniques.

Purpose of the Study:

• To develop a generative model for multi-atlas segmentation that is independent of training image intensities.
• To address limitations of existing methods in handling cross-modality and cross-dataset MRI segmentation.

Main Methods:

• A probabilistic generative model is proposed, exploiting within-region voxel intensity consistency in the target volume.
• Variational expectation-maximization (EM) algorithm is used to derive the most likely segmentation.
• The model was tested using T1-weighted MRI atlases for segmenting proton-density (PD) weighted brain MRI scans.

Main Results:

• The generative model successfully segmented PD-weighted brain MRI scans using T1-weighted atlases, a task challenging for traditional methods.
• The proposed method demonstrated significant improvement over established techniques like majority voting and STAPLE.
• This approach effectively handles voxel intensity inconsistencies between atlases and target images.

Conclusions:

• The generative model offers a robust solution for multi-atlas segmentation, overcoming intensity variations.
• This method expands the applicability of multi-atlas segmentation to diverse imaging modalities and datasets.
• The probabilistic framework provides accurate and reliable brain MRI segmentation.