GraformerDIR: Graph convolution transformer for deformable image registration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deformable image registration (DIR) is crucial for disease diagnosis.
• Transformer models improve DIR by capturing long-range dependencies, overcoming convolution limitations.
• However, Transformer-based DIR struggles with high-quality long-range dependency extraction due to connection issues.

Purpose of the Study:

• To propose GraformerDIR, a novel deformable image registration framework.
• To enhance the extraction of high-quality long-range dependencies in DIR using graph convolution Transformers.

Main Methods:

• Introduced a Graformer layer combining a Graformer module for long-range dependencies and a Chebyshev graph convolution module to expand the receptive field.
• Integrated the Graformer layer into the feature extraction network of a DIR framework, creating GraformerDIR.
• Evaluated GraformerDIR on brain datasets (OASIS, LPBA40, MGH10) and a cardiac dataset (ACDC).

Main Results:

• GraformerDIR demonstrated significant performance improvements over VoxelMorph on the OASIS dataset: 4.6% increase in Dice Similarity Coefficient (DSC) and 0.055 mm reduction in Average Symmetric Surface Distance (ASD).
• Reduced the non-positive Jacobian determinant rate by approximately 60 times on the OASIS dataset.
• Showcased strong generalizability on the unseen MGH10 dataset with 4.1% DSC and 0.084 mm ASD improvements.
• Achieved promising results on the ACDC cardiac dataset, indicating clinical practicability.

Conclusions:

• GraformerDIR leverages Transformer and graph convolution advantages for superior DIR performance.
• The proposed framework achieves comparable results to state-of-the-art methods like VoxelMorph.
• GraformerDIR offers enhanced accuracy and generalizability for medical image registration.