Lightweight cross-resolution coarse-to-fine network for efficient deformable medical image registration

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Anatomy

Background:

• Accurate and efficient deformable medical image registration is critical.
• Deep learning methods offer high accuracy but often lack efficiency due to large model sizes and slow inference.
• Existing methods struggle to balance accuracy and computational efficiency.

Purpose of the Study:

• To develop a lightweight yet accurate deformable medical image registration framework.
• Introduce LightCRCF, a novel framework addressing the accuracy-efficiency trade-off in medical image registration.

Main Methods:

• Utilizes an ultra-lightweight U-Net architecture (0.1M parameters) for efficiency.
• Employs a cross-resolution coarse-to-fine (C2F) strategy to progressively decompose deformation fields.
• Incorporates Texture-aware Reparameterization (TaRep) and Group-flow Reparameterization (GfRep) modules for enhanced accuracy.
• Applies structural reparameterization to maintain training accuracy and efficient inference.

Main Results:

• LightCRCF achieves accuracy comparable to state-of-the-art methods on multiple MRI and CT datasets.
• Demonstrates significantly superior performance in efficiency metrics (parameters, VRAM, FLOPs, inference time).
• Outperforms other efficiency-focused methods in accuracy.

Conclusions:

• LightCRCF presents a favorable balance between accuracy and efficiency in deformable medical image registration.
• The framework shows potential for practical clinical applications.
• Open-source code is available for reproducibility and further development.