Fourier-Net+: Band-Limited Spatial Representation for Efficient Medical Image Registration

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• U-Net networks are standard for unsupervised image registration, predicting dense displacement fields.
• Processing high-resolution volumetric data with these networks is resource-intensive and slow.

Purpose of the Study:

• Introduce Fourier-Net+, an efficient image-domain deformable registration framework.
• Enhance registration performance and encourage diffeomorphism with cascaded and diffeomorphic variants.

Main Methods:

• Fourier-Net+ utilizes deterministic Fourier-domain band-limiting for efficient down/up-sampling.
• A parameter-free decoder learns a band-limited representation of the displacement field.
• All network layers are real-valued for broad modality compatibility.

Main Results:

• Fourier-Net+ variants achieve comparable results to state-of-the-art methods across five diverse datasets.
• Demonstrated faster training/inference, lower memory footprint, and reduced mult-adds.
• Diff-Fourier-Net+ significantly outperformed baselines on 3-D-CMR data in Dice, HD, and clinical metrics.

Conclusions:

• Fourier-Net+ provides an efficient and effective solution for deformable image registration.
• The framework's efficiency enables large-scale 3-D registration training on low-VRAM GPUs.
• Proposed variants enhance performance and diffeomorphism, making it suitable for various medical imaging applications.