Inverse Consistency by Construction for Multistep Deep Registration

Area of Science:

• Medical imaging
• Computer vision
• Computational anatomy

Background:

• Image registration aligns medical images, crucial for diagnosis and treatment planning.
• Inverse consistency ensures that the transformation from image A to B is the exact inverse of the transformation from B to A, a vital property for accurate registration.
• Current neural network approaches for image registration often lack guaranteed inverse consistency.

Purpose of the Study:

• To introduce a novel neural network architecture that inherently enforces inverse consistency in image registration.
• To extend this technique for multi-step and coarse-to-fine registration frameworks.
• To validate the effectiveness of the proposed method in achieving accurate and inverse-consistent image registration.

Main Methods:

• A new neural network design is proposed that parameterizes transformations using Lie groups, ensuring inverse consistency by construction.
• The technique is extended to multi-step registration by composing networks while preserving inverse consistency.
• The method is applied to both synthetic 2D data and four 3D medical imaging datasets.

Main Results:

• The proposed technique successfully generates inverse-consistent image registrations.
• Excellent registration accuracy was achieved across various datasets, including synthetic and real 3D medical images.
• The multi-step approach enabled effective coarse-to-fine inverse-consistent registration.

Conclusions:

• The proposed neural network structure provides a simple yet effective method for achieving inverse-consistent image registration.
• This approach guarantees inverse consistency by design, eliminating the need for post-hoc correction.
• The technique demonstrates significant potential for improving the reliability and accuracy of medical image registration applications.