One-Shot Learning for Deformable Medical Image Registration and Periodic Motion Tracking

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deformable image registration is crucial in medical imaging, with deep learning showing promise.
• Deep learning methods often require large datasets and struggle with unseen data.
• One-shot learning offers a solution by reducing the need for extensive training data.

Purpose of the Study:

• To present a one-shot learning approach for periodic motion tracking in 3D and 4D medical datasets.
• To address the limitations of traditional deep learning methods in deformable registration.

Main Methods:

• Employed a U-Net architecture combined with a coarse-to-fine strategy.
• Integrated a differential spatial transformer module for registration.
• Developed a method to calculate the inverse registration vector field simultaneously for 3D datasets.

Main Results:

• The one-shot registration approach demonstrated effectiveness in tracking periodic motion.
• Achieved competitive registration accuracy on publicly available 3D and 4D datasets.
• Validated the algorithm's ability to handle unseen image data.

Conclusions:

• The proposed one-shot registration method is a viable alternative for 3D and 4D motion tracking.
• Applicable as a standalone tool or in early study phases before large dataset availability.
• Offers a robust solution for periodic motion tracking in medical imaging.