Dlung: Unsupervised Few-Shot Diffeomorphic Respiratory Motion Modeling

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Lung image registration is crucial for applications like respiratory motion modeling.
• Unsupervised methods are attractive but struggle with limited data and diffeomorphic properties.
• Existing methods often fail to preserve lung topology during large deformations.

Purpose of the Study:

• To develop an unsupervised, few-shot learning-based method for diffeomorphic lung image registration.
• To address the challenges of limited data and topology preservation in lung image analysis.
• To improve the accuracy and efficiency of respiratory motion modeling.

Main Methods:

• Introduced Dlung, an unsupervised few-shot learning framework for lung image registration.
• Utilized fine-tuning techniques to overcome data limitations.
• Employed the scaling and squaring method to ensure diffeomorphic (topology-preserving) registration.
• Performed atlas-based registration on 4D spatio-temporal lung images.

Main Results:

• Dlung achieved the highest accuracy among compared methods.
• The registration process successfully preserved diffeomorphic properties.
• The method demonstrated effectiveness in constructing accurate respiratory motion models with limited data.
• Fast and accurate motion modeling was achieved.

Conclusions:

• Dlung offers a robust solution for unsupervised, few-shot diffeomorphic lung image registration.
• The developed method enhances respiratory motion modeling capabilities, especially in data-scarce scenarios.
• This research advances the field of lung image analysis and motion modeling.