Structure-Aware DeformConv and Hierarchical Multi-scale Transformer for Medical Image Registration

Area of Science:

• Medical image analysis
• Computer vision
• Biomedical engineering

Background:

• Multimodal medical image registration is vital for diagnosis and treatment planning.
• Challenges include complex spatial deformations and cross-modal feature matching difficulties.
• Existing methods struggle with accuracy and robustness across diverse modalities.

Purpose of the Study:

• To propose an unsupervised and discriminator-free multimodal registration method, SA-HMT.
• To enhance cross-modal feature matching for improved registration accuracy.
• To address the limitations of current methods in handling complex spatial deformations.

Main Methods:

• Developed a multi-scale skip Transformer module with a hierarchical architecture for multi-scale deformation feature capture.
• Utilized a structure-aware deformable convolution module with a two-stage joint mechanism for enhanced feature matching.
• Employed a dual loss function for unsupervised learning.

Main Results:

• SA-HMT demonstrated effectiveness across five public datasets (CT-MR, lung CT, brain CT-MR, cardiac MRI) and clinical abdominal data.
• Achieved improved performance in core indicators like Dice Similarity Coefficient (DSC) compared to the advanced R2Net method.
• Registration accuracy was generally comparable or superior to existing state-of-the-art approaches.

Conclusions:

• SA-HMT offers a robust and accurate solution for unsupervised multimodal medical image registration.
• The proposed modules effectively address cross-modal feature matching challenges.
• This method holds significant potential for advancing medical image analysis in clinical practice.