ORRN: An ODE-Based Recursive Registration Network for Deformable Respiratory Motion Estimation With Lung 4DCT Images

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning

Background:

• Deformable Image Registration (DIR) is crucial for quantifying medical data deformation.
• Deep learning methods have improved DIR accuracy and speed for image pairs.
• Existing methods struggle with 4D data's complex organ motion (e.g., respiration, heartbeat).

Purpose of the Study:

• To present ORRN, an Ordinary Differential Equations (ODE)-based recursive network for 4D medical image registration.
• To effectively model time-varying organ motion in 4D datasets.
• To improve registration accuracy and efficiency for dynamic medical data.

Main Methods:

• ORRN estimates time-varying voxel velocities to model 4D deformation using an ODE.
• A recursive registration strategy integrates voxel velocities to progressively estimate deformation fields.
• The method was evaluated on two public 4DCT lung datasets (DIRLab, CREATIS).

Main Results:

• ORRN achieved superior performance in 3D+t deformation tracking and exhale-to-inhale registration.
• The method yielded a minimal Target Registration Error (1.24 mm and 1.26 mm).
• Results showed high deformation plausibility (<0.001% folding) and computational efficiency (<1s per volume).

Conclusions:

• ORRN demonstrates significant accuracy, plausibility, and efficiency in 4D image registration.
• The method is effective for both group-wise and pair-wise registration tasks.
• ORRN has implications for respiratory motion estimation in radiation therapy and robotic surgery planning.