GraphRegNet: Deep Graph Regularisation Networks on Sparse Keypoints for Dense Registration of 3D Lung CTs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning methods show promise in medical image registration.
• Current U-Net-like architectures struggle with large 3D deformations, limiting accuracy in tasks like lung CT and abdominal MRI registration.
• State-of-the-art deep learning registration accuracy still lags behind conventional frameworks.

Purpose of the Study:

• To develop a novel deep learning approach, GraphRegNet, for accurate dense deformable medical image registration, specifically addressing the challenge of large 3D deformations.
• To improve registration accuracy and handle memory constraints in 3D medical volumes.

Main Methods:

• GraphRegNet, a sparse keypoint-based geometric network, formulates registration as predicting displacement vectors on a sparse grid.
• Combines convolutional and graph neural network layers for efficient displacement regularization.
• Leverages discrete dense displacement maps, inspired by 2D optical flow methods.

Main Results:

• GraphRegNet demonstrates substantial improvements in exhale to inhale lung CT registration.
• Achieved high accuracy with a TRE (Target Registration Error) below 1.4 mm.
• Outperforms other deep learning methods in experimental evaluations.

Conclusions:

• GraphRegNet offers an effective solution for dense deformable medical image registration, especially for large 3D deformations.
• The proposed method achieves state-of-the-art accuracy, surpassing existing deep learning techniques.
• Publicly available code facilitates further research and application.