Gaussian primitives for deformable image registration.

Area of Science:

• Medical imaging
• Computational anatomy
• Radiotherapy physics

Background:

• Deformable image registration (DIR) is crucial for radiotherapy, compensating for anatomical changes.
• Existing DIR methods face challenges with computational efficiency and generalization.
• There is a need for improved DIR techniques that are both fast and accurate.

Purpose of the Study:

• To develop a novel optimization-based DIR method, GaussianDIR.
• To reduce computational overhead in DIR.
• To preserve the generalization capabilities of iterative methods while enhancing interpretability.

Main Methods:

• Proposed GaussianDIR, a framework using adaptive Gaussian primitives to represent deformation fields.
• Each primitive is defined by its center, covariance, and local rigid deformation.
• Voxel displacements are calculated by blending deformations from neighboring primitives.

Main Results:

• GaussianDIR achieved a target registration error of 1.00±1.11 mm in 2.5 s on the DIRLab lung dataset.
• Improved Dice Similarity Coefficient (DSC) on OASIS brain and ACDC cardiac datasets compared to state-of-the-art methods.
• Outperformed a data-driven method by 6.3% in DSC on the IXI dataset, demonstrating superior generalization.

Conclusions:

• GaussianDIR integrates high registration accuracy with computational efficiency and interpretability.
• The method challenges the notion of slow iterative registration and overcomes generalization limitations of data-driven approaches.
• GaussianDIR shows potential for real-time clinical applications in radiotherapy.