4D-CT deformable image registration using unsupervised recursive cascaded full-resolution residual networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image registration is crucial for diagnosing and treating diseases.
• Four-dimensional computed tomography (4D-CT) generates dynamic volumetric data, posing registration challenges.
• Existing registration methods often require extensive training data or lack robustness.

Purpose of the Study:

• To introduce a novel unsupervised deep learning network for abdominal 4D-CT image registration.
• To evaluate the performance and generalization capability of the proposed method against existing techniques.

Main Methods:

• A recursive cascaded full-resolution residual network (RCFRR-Net) was developed for end-to-end unsupervised learning.
• The network comprises three cascaded subnetworks trained jointly using image similarity and deformation regularization losses.
• Testing was performed on internal 4D-CT, public DIRLAB 4D-CT, and 4D cone-beam CT (4D-CBCT) datasets.

Main Results:

• RCFRR-Net demonstrated consistent and significant performance gains compared to the demon method, VoxelMorph, and a recursive cascaded network.
• The proposed method achieved superior accuracy and generalization capability across diverse medical imaging datasets.
• Unsupervised training eliminated the need for ground truth deformation fields during the learning process.

Conclusions:

• RCFRR-Net offers a robust and accurate solution for 4D-CT image registration.
• The unsupervised, end-to-end deep learning approach shows significant potential for clinical applications.
• This method advances the field of medical image registration through improved performance and generalization.