Few-shot learning for deformable image registration in 4DCT images.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Online adaptive radiotherapy requires rapid and accurate 4D deformable image registration (DIR).
• Traditional DIR methods can be computationally intensive and suffer from reference image selection bias.

Purpose of the Study:

• To develop a rapid, accurate, and computationally efficient 4D DIR approach for online adaptive radiotherapy.
• To mitigate reference image bias in 4D DIR using a few-shot learning strategy.

Main Methods:

• A deep learning (DL)-based few-shot registration network (FR-Net) was developed.
• FR-Net generates deformation vector fields from each respiratory phase to an implicit reference image.
• The network was pretrained on unlabeled 4D data and optimized for intensity similarity on specific 4DCT scans.

Main Results:

• FR-Net achieved an average target registration error of 1.48 mm and 1.16 mm on DIR-Lab and POPI datasets, respectively.
• Optimization of one 4DCT scan required approximately 2 minutes.
• The method demonstrated superior registration accuracy and low computational time compared to state-of-the-art techniques.

Conclusions:

• A few-shot groupwise DIR algorithm for 4DCT images was successfully developed.
• The FR-Net approach shows promising registration performance and computational efficiency for online adaptive radiotherapy.
• This work highlights the potential of DL models combined with groupwise registration and few-shot learning for efficient 4DCT registration.