F3RNet: full-resolution residual registration network for deformable image registration

Area of Science:

• Medical imaging
• Deep learning
• Image registration

Background:

• Deformable image registration (DIR) is vital for image-guided therapies.
• Deep learning (DL) has shown success in DIR, but struggles with accurate local alignment in severely deformed regions.
• Existing DL methods often overlook precise alignment of hard-to-register areas, impacting surgical targeting.

Purpose of the Study:

• To develop a novel unsupervised deep learning network for accurate deformable registration of severely deformed organs.
• To address the limitations of current DIR methods in achieving precise local alignments.
• To improve the sensitivity of registration algorithms to challenging anatomical variations.

Main Methods:

• Proposed a full-resolution residual registration network (F3RNet) utilizing two parallel processing streams.
• One stream leverages full-resolution information for accurate voxel-level registration.
• The other stream learns multi-scale residual representations for robust recognition, with factorized 3D convolution for efficiency.

Main Results:

• Validated F3RNet on intra-patient abdominal CT-MRI and public thorax CT datasets.
• Demonstrated promising results in both multimodal and unimodal registration compared to state-of-the-art methods.
• Achieved accurate overall and local registration, crucial for clinical applications.

Conclusions:

• F3RNet effectively combines high-resolution and multi-scale information for superior registration accuracy.
• The network achieves rapid image registration, with a run time under 3 seconds per pair on GPU.
• Future work will focus on optimizing high-resolution information processing and multi-scale representation fusion.