AtlasMorph: Learning conditional deformable templates for brain MRI.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning

Background:

• Deformable templates (atlases) with probabilistic labels are crucial for medical image analysis.
• Developing these templates is computationally intensive, leading to limited availability and suboptimal choices for population studies.
• Existing templates may not accurately represent populations with significant anatomical variations.

Purpose of the Study:

• To develop an efficient machine learning framework for generating subject-specific, conditional templates.
• To leverage subject attributes (e.g., age, sex) and segmentations for template creation.
• To improve the accuracy and representativeness of anatomical templates in medical imaging.

Main Methods:

• Utilized convolutional registration neural networks to learn a template-generating function.
• Incorporated subject-specific attributes (age, sex) to condition template output.
• Leveraged available segmentations to generate probabilistic anatomical label maps for templates.
• Demonstrated the method on 3D brain MRI datasets.

Main Results:

• Successfully learned high-quality, representative templates for diverse populations.
• Demonstrated that annotated conditional templates significantly enhance registration accuracy compared to unlabeled or unconditional templates.
• Showcased superior performance over traditional template construction methods.

Conclusions:

• The proposed machine learning framework efficiently generates representative, conditional anatomical templates.
• Conditional templates improve registration accuracy and overcome limitations of static, population-averaged atlases.
• This approach offers a powerful tool for computational anatomy and population-based medical image analysis.