Patch spaces and fusion strategies in patch-based label fusion

Area of Science:

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

Background:

• Patch-based methods are effective for multi-atlas segmentation of biomedical images.
• Current methods often warp atlases to target space, introducing interpolation errors.
• Registration is crucial for anatomical priors but can be computationally intensive.

Purpose of the Study:

• To investigate extracting information from native atlas and target image spaces to avoid interpolation artifacts.
• To unify similarity-based and learning-based label fusion strategies.
• To enhance the accuracy and efficiency of multi-atlas segmentation.

Main Methods:

• Developed a framework extracting appearance and label information from native image spaces.
• Formulated a common approach for similarity-based and learning-based label fusion.
• Evaluated the framework on subcortical brain structures and fetal brain MRI.

Main Results:

• Extracting patches from native atlas space significantly outperformed warping atlases to target space.
• Learning-based fusion generally showed superior performance compared to similarity-based methods.
• Using native space patches reduced computational demands for the learning-based approach.

Conclusions:

• Combining learning-based label fusion with native atlas patches offers the best performance.
• This combined approach reduces computational time compared to conventional similarity-based methods.
• Native space patch extraction is a key strategy for improving multi-atlas segmentation accuracy and efficiency.