GraphRegNet

GraphRegNet predicts displacement vectors on a sparse irregular grid of keypoints to achieve dense registration of 3D lung CT images and to handle large 3D deformations for accurate medical image registration.

Key Features:

• Deep graph regularization network: Combines convolutional neural network layers with graph neural network layers within a unified architecture for displacement regularization.
• Sparse keypoint formulation: Predicts displacement vectors on a sparse irregular grid of distinctive keypoints rather than dense voxel-wise displacement maps.
• Dense deformation field reconstruction: Produces well-regularized dense registration fields for 3D lung CT volumes from sparse keypoint predictions.
• Memory efficiency: Reduces memory requirements compared to traditional dense displacement maps used in 3D medical volumes.
• Large-deformation handling: Targets large 3D deformations such as inhale-to-exhale lung CT and interpatient abdominal MRI registration.
• Learning-based registration: Supports both supervised and unsupervised medical image registration approaches.
• Fast inference: Enables near real-time applications through learning-based fast inference.
• Performance metric: Demonstrated target registration error (TRE) below 1.4 mm on exhale-to-inhale lung CT experiments.
• Optical flow inspiration: Adapts concepts from FlowNet and PWC-Net, including the use of discrete dense displacement maps, to 3D medical imaging.

Scientific Applications:

• Exhale-to-inhale lung CT registration: Registration of inhale–exhale lung CT pairs with sub-1.4 mm TRE for tracking anatomical correspondence.
• Interpatient abdominal MRI registration: Alignment of abdominal MRI volumes across patients to accommodate large intersubject anatomical variability.
• General 3D medical image registration: Supervised and unsupervised registration tasks requiring accurate dense deformation fields and large-deformation handling.

Methodology:

Formulates registration as prediction of displacement vectors on a sparse irregular grid of keypoints and regularizes these displacements via a unified CNN+GNN architecture; adapts discrete dense displacement map concepts from FlowNet and PWC-Net to 3D medical imaging and supports supervised and unsupervised learning-based training.