Wave-Reg: full-stage wavelet-guided image registration framework with cross-scale correction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image registration is crucial but challenged by complex deformations.
• Deep learning methods face issues like information loss and the "small objects move fast" problem.
• Existing coarse-to-fine architectures accumulate deformation errors.

Purpose of the Study:

• To introduce Wave-Reg, a novel spatial-frequency registration framework.
• To address information loss, the "small objects move fast" problem, and accumulated errors in medical image registration.
• To improve registration accuracy for large deformation and multi-modality tasks.

Main Methods:

• Developed Wave-Reg using a wavelet pyramid architecture with discrete wavelet transform (DWT).
• Employed DWT-guided ConvNet for feature extraction to minimize detail loss.
• Utilized inverse DWT-guided Swin Transformer for deformation reconstruction, mitigating the "small objects move fast" problem.
• Integrated a cross-scale self-correction module with Heun's predictor-corrector method to refine deformation fields.

Main Results:

• Demonstrated substantial gains in registration accuracy across three datasets.
• Achieved superior performance in both large deformation and multi-modality registration tasks.
• Validated the effectiveness of spatial-frequency feature learning and predictor-corrector refinement.

Conclusions:

• Wave-Reg offers an effective solution to longstanding challenges in medical image registration.
• Spatial-frequency learning and predictor-corrector refinement are key to improving registration accuracy.
• The proposed framework shows significant potential for advancing medical image analysis.