Domain-Shift Immunity in Deep Registration
Paper#Image Registration🔬 Research|Analyzed: Jan 3, 2026 19:10•
Published: Dec 29, 2025 02:10
•1 min read
•ArXivAnalysis
This paper challenges the common belief that deep learning models for deformable image registration are highly susceptible to domain shift. It argues that the use of local feature representations, rather than global appearance, is the key to robustness. The authors introduce a framework, UniReg, to demonstrate this and analyze the source of failures in conventional models.
Key Takeaways
- •Deep deformable registration models can be inherently robust to domain shift.
- •Local feature consistency is a key driver of robustness.
- •Dataset-induced biases in early convolutional layers can cause failures under modality shift.
- •UniReg framework demonstrates domain-shift immunity using fixed, pre-trained feature extractors.
Reference / Citation
View Original"UniReg exhibits robust cross-domain and multi-modal performance comparable to optimization-based methods."