Geometric Algebra-based Dual-branch Graph Convolutional Network for Deformable Medical Image Registration with...

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning in Healthcare

Background:

• Deformable medical image registration (DMIR) is crucial for medical image analysis.
• Graph Neural Networks (GCNs) show promise but struggle with feature similarity in deep, multi-stage models, hindering accuracy for large deformations.
• Existing methods face challenges in capturing fine-grained structural differences and efficient registration.

Purpose of the Study:

• To introduce a novel method for accurate and efficient deformable medical image registration, particularly for large volumetric deformations.
• To overcome limitations of GCNs in feature extraction and maintain accuracy in complex registration tasks.
• To enhance the capture and integration of image features through global contextual interactions.

Main Methods:

• Introduced Geometric Algebra-based MorphNet (GAMN) to leverage geometric algebra for restructuring information in non-Euclidean space.
• Implemented a Bi-Directional Flow (BDF) mechanism for progressive deformation field prediction using high-level features.
• Employed Specified Skip Propagation (SSP) for efficient, adaptive feature transfer across network levels.

Main Results:

• Achieved significant improvements in Dice similarity scores: 0.7% on OASIS and 1.4% on LPBA40 datasets compared to the PIViT baseline.
• Qualitative and quantitative analyses demonstrated enhanced registration accuracy and efficiency.
• Ablation studies confirmed the positive impact of each module (GAMN, BDF, SSP) on network performance.

Conclusions:

• The proposed GAMN and BDF method, enhanced by SSP, effectively addresses challenges in deformable medical image registration.
• This approach improves accuracy and efficiency, especially for large volumetric deformations, outperforming existing methods.
• The findings highlight the potential of geometric algebra and advanced flow mechanisms in advancing medical image analysis.