CycleMorph: Cycle consistent unsupervised deformable image registration.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep learning methods offer fast and effective medical image registration.
• Existing methods struggle with preserving topology during deformation.
• Registration vector fields are crucial for accurate alignment.

Purpose of the Study:

• Introduce CycleMorph, a cycle-consistent deep learning model for deformable image registration.
• Address the limitation of topology preservation in current deep learning registration techniques.
• Provide a flexible and efficient registration method applicable to diverse datasets.

Main Methods:

• Developed a cycle-consistent adversarial network for deformable image registration.
• Incorporated cycle consistency as an implicit regularization for topology preservation.
• Designed a flexible architecture applicable to 2D and 3D registration problems.
• Enabled multi-scale implementation to manage memory constraints in large volume registration.

Main Results:

• CycleMorph achieves effective and accurate registration on diverse medical and non-medical image pairs.
• The method demonstrates ultra-fast computational times, completing registration in seconds.
• Qualitative and quantitative evaluations confirm the effectiveness of cycle consistency in preserving topology.
• The approach shows comparable performance to state-of-the-art classical methods.

Conclusions:

• CycleMorph significantly improves deformable image registration by ensuring topological integrity.
• The proposed method offers a fast, accurate, and flexible solution for 2D and 3D image registration tasks.
• Cycle consistency serves as a powerful regularization technique for deep learning-based medical image analysis.