GroupRegNet

GroupRegNet performs groupwise one-shot deep learning-based four-dimensional (4D; three spatial dimensions plus time) image registration to compute deformation fields that align dynamic medical images for accurate temporal and spatial registration.

Key Features:

• One-Shot Learning Approach: Employs a one-shot learning strategy to generate deformation fields that align all images within a group to a common template.
• Implicit Template Utilization: Leverages an implicit template to minimize errors related to reference selection across multiple images.
• Simplified Network Design: Uses a streamlined network architecture that avoids dividing input images into patches.
• Efficiency and Accuracy: Reduces inference time while maintaining high registration accuracy comparable to top conventional methods.

Scientific Applications:

• 4D Medical Image Registration: Aligns temporal and spatial information in dynamic medical imaging datasets for clinical and research use.
• Respiratory-Binned 4D Computed Tomography Analysis: Handles respiratory-binned 4D-CT datasets to register complex deformations across time.
• Benchmarking and Method Comparison: Quantitative assessments on two public respiratory-binned 4D-computed tomography datasets show it outperforms existing deep learning-based methods and achieves accuracy comparable to the conventional method pTVreg.

Methodology:

Employs a convolutional neural network (CNN) that replaces traditional motion models and parameters and optimizes network weights to achieve image alignment; uses one-shot learning to produce deformation fields that align all group images to an implicit template; the network design omits patch-based input division and its optimization mirrors conventional iterative optimization approaches.