ElasticMorph: Plug-and-play second-order elastic regularization for medical image registration.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Biomedical Engineering

Background:

• Deformable image registration is vital for clinical applications.
• Current learning-based methods use simple smoothness losses, leading to physically implausible deformations like tissue folding.
• This creates a trade-off between registration accuracy and anatomical correctness.

Purpose of the Study:

• To develop and validate a novel, physics-informed regularizer for deformable image registration.
• The regularizer aims to simultaneously improve registration accuracy and physical plausibility.
• Achieve this with negligible computational overhead.

Main Methods:

• Proposed ElasticMorph, a plug-and-play loss function based on the Navier-Cauchy equation.
• The regularizer penalizes curvature and divergence in deformation fields.
• Evaluated on brain MRI datasets (IXI, LPBA-40) integrated into CNN and transformer models, quantifying performance with Dice Similarity Coefficient (DSC) and percentage of negative Jacobians (%negJ).

Main Results:

• ElasticMorph consistently improved registration accuracy, increasing mean DSC by 0.56-4.92%.
• Reduced folding artifacts (%negJ) by 24%-42% across tested models and datasets.
• Achieved with minimal training time increase (≤8%) and memory usage (≤15%), with no inference cost.

Conclusions:

• Enforcing second-order linear elasticity is a robust and efficient strategy for deformable registration.
• ElasticMorph offers a practical, principled, and plug-and-play solution for more accurate and physically plausible registration.
• This holds significant potential for improving the reliability of biomedical image analysis tasks.