Deformable image registration using multi-resolution vision Transformer for cardiac motion estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Deformable registration is vital for cardiac magnetic resonance (CMR) imaging analysis.
• Challenges include intensity inhomogeneity and complex cardiac deformations.

Purpose of the Study:

• To develop a novel CNN-Transformer framework for accurate deformable registration of CMR images.
• To improve cardiac motion estimation for better diagnosis and treatment of heart diseases.

Main Methods:

• A convolutional projection Transformer block was designed for efficient self-attention and long-range spatial correspondence modeling.
• A cooperative learning pattern fused global and local features.
• A multi-resolution strategy optimized model parameters in a coarse-to-fine manner.

Main Results:

• The proposed method demonstrated superior performance on three CMR datasets for intra-subject registration.
• Achieved better Dice overlap and lower surface distance compared to existing methods.
• Outperformed four non-learning-based and three deep-learning-based methods.

Conclusions:

• The novel CNN-Transformer framework offers superior accuracy and lower complexity for CMR image registration.
• This method facilitates more precise cardiac motion estimation for clinical assessments.
• The approach enhances the diagnosis and treatment of cardiac diseases.