Masked Deformation Modeling for Volumetric Brain MRI Self-Supervised Pre-Training.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in neuroscience
• Machine learning for healthcare

Background:

• Self-supervised learning (SSL) shows promise for medical image segmentation but often lacks brain-specific priors.
• Existing SSL methods pre-trained on non-brain data underperform on detailed brain parcellation tasks.

Purpose of the Study:

• To introduce Masked Deformation Modeling (MDM), a novel SSL strategy tailored for human brain MRI segmentation.
• To enhance the performance of brain MRI segmentation by incorporating brain-specific priors into SSL.

Main Methods:

• MDM employs atlas-guided patch sampling from individual brain MRIs and an MNI152 template.
• Random feature-aligned masking is applied to sampled volumes before feature extraction via a U-Net.
• Latent features are decoded by intensity and deformation field heads to restore volumes and predict spatial transformations.

Main Results:

• MDM demonstrated superior performance across multiple brain segmentation datasets (parcellation, lesion, tumor).
• The method significantly outperformed existing state-of-the-art medical SSL approaches.
• MDM achieved a reduction in annotation effort by at least 40%.

Conclusions:

• MDM effectively addresses the limitations of general SSL methods in brain MRI segmentation.
• The proposed strategy offers a powerful tool for improving the accuracy and efficiency of brain parcellation and segmentation.
• MDM shows potential to substantially decrease the manual annotation burden in medical imaging research.