Feature-centric registration of large deformed images using transformers and correlation distance

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deformable medical image registration requires robust networks and similarity metrics.
• Handling large deformations and lack of ground truth data are key challenges.

Purpose of the Study:

• To develop a robust registration network and feature-based loss for large deformations without ground truth.
• To improve accuracy in medical image registration tasks.

Main Methods:

• Implemented a coarse-to-fine displacement vector field (DVF) estimation.
• Integrated Transformer's feature attention mechanism.
• Proposed a novel feature correlation-based distance metric using symmetric correlation matrices and network-extracted features.

Main Results:

• The feature correlation-based loss effectively achieves accurate registration without ground truth data.
• Demonstrated success in mono-modality abdomen CT registration and brain MRI atlas registration.
• Showed improvements in Dice similarity coefficient and other evaluation metrics.

Conclusions:

• The proposed method offers a robust solution for deformable medical image registration, especially in scenarios with large deformations and missing ground truth.
• The feature correlation-based loss function is a valuable advancement for medical image analysis.