An ODE-based multi-resolution parallel network for respiratory motion estimation.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Respiratory motion significantly displaces lesions during lung puncture procedures.
• Accurate pulmonary motion estimation is crucial for surgical guidance.
• Complex lung deformations and internal structures challenge current motion estimation techniques.

Purpose of the Study:

• To develop an advanced method for estimating pulmonary respiratory motion.
• To improve the accuracy and reliability of lung motion estimation for surgical applications.
• To address limitations in current deep learning models for 4D CT data.

Main Methods:

• Proposed a multi-resolution parallel network architecture incorporating neural ordinary differential equations (neural ODE).
• Utilized neural ODE to explicitly model temporal continuity in 4DCT data, ensuring realistic deformations.
• Employed a multi-resolution parallel structure for recursive feature refinement to enhance prediction capabilities.

Main Results:

• The proposed method demonstrated superior performance compared to existing deep learning approaches.
• Achieved consistently high accuracy in lung motion estimation across all respiratory phases.
• Generated transformations that better align with physiological respiratory motion patterns.

Conclusions:

• The novel neural ODE-based network effectively estimates pulmonary respiratory motion with high accuracy.
• This method offers improved guidance for surgical interventions involving lung lesions.
• The approach enhances feature representation and prediction, leading to superior registration accuracy in 4DCT analysis.