Coordinate Translator for Learning Deformable Medical Image Registration

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning (DL) methods for deformable image registration commonly use convolutional neural networks (CNNs).
• CNNs in registration must simultaneously extract image features and understand spatial coordinate systems, which is challenging.
• This dual requirement can limit the performance of traditional CNNs in registration tasks.

Purpose of the Study:

• To develop a novel DL approach for deformable image registration that addresses the limitations of traditional CNNs.
• To improve the accuracy and efficiency of unsupervised 3D magnetic resonance image registration.

Main Methods:

• Introduced Coordinate Translator, a differentiable module for identifying feature correspondences and their coordinates without training.
• Proposed im2grid, a novel deformable registration network utilizing multiple Coordinate Translators with CNN hierarchical features.
• Implemented a coarse-to-fine strategy for outputting deformation fields.

Main Results:

• The im2grid network demonstrated superior performance compared to state-of-the-art DL and non-DL methods.
• Qualitative and quantitative experiments on unsupervised 3D magnetic resonance image registration confirmed im2grid's effectiveness.
• The Coordinate Translator module successfully offloaded the burden of coordinate system understanding from the CNNs.

Conclusions:

• The proposed im2grid network, enhanced by the Coordinate Translator, represents a significant advancement in DL-based deformable image registration.
• This approach enables CNNs to focus more effectively on feature extraction, leading to improved registration accuracy.
• im2grid offers a promising solution for accurate and efficient 3D medical image registration.