PC-Reg: A pyramidal prediction-correction approach for large deformation image registration

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Deep Learning in Medical Imaging

Background:

• Deformable image registration is crucial for medical image analysis.
• Current deep learning methods (e.g., VoxelMorph, TransMorph) excel at small deformations but struggle with large ones.
• Large deformations pose significant challenges for accurate medical image alignment.

Purpose of the Study:

• To introduce PC-Reg, a novel pyramidal Prediction and Correction method for deformable image registration.
• To address the limitations of existing methods in handling large deformations.
• To improve the accuracy and robustness of medical image registration across various deformation scales.

Main Methods:

• PC-Reg employs a multi-scale, pyramidal approach treating registration as an ordinary differential equation (ODE) across scales.
• It utilizes a predictor-corrector strategy, iteratively refining the deformation vector field with residual and correction flows.
• A distillation loss is incorporated to transfer high-resolution details to lower scales, enhancing feature preservation.

Main Results:

• PC-Reg demonstrates significant improvements in registration accuracy for both large and small deformations in inter-patient scenarios.
• The method effectively mitigates issues related to multiscale accumulated upsampling and diffeomorphic integration errors.
• Experimental results validate the efficacy of the proposed pyramidal prediction and correction framework.

Conclusions:

• PC-Reg offers a robust and effective solution for deformable medical image registration, particularly excelling with large deformations.
• The ODE-inspired, multi-scale approach enhances registration accuracy and stability.
• The method shows promise for advancing medical image analysis tasks requiring precise anatomical alignment.