Deformable image registration is a fundamental challenge in medical image analysis. Recent deep‑learning approaches fall into two categories: one‑pass architectures that complete registration in a single forward pass, and recursive architectures that refine the result through multiple iterations. While recursion can improve accuracy, it incurs considerable computational cost and slower inference.
This work questions whether recursion is truly required for high‑performance registration and introduces M2CG, a one‑pass network equipped with Multi‑level Multi‑space Contextual Guidance. M2CG performs multi‑step progressive pyramid registration within a single forward pass; each step is guided by contextual cues from three levels (previous, current, subsequent) and four spaces (image space, image feature space, segmentation space, segmentation feature space). The model supports both unsupervised training (no labels) and semi‑supervised training (using segmentation labels for a subset of data).
Evaluated on two inter‑/intra‑patient benchmark tasks across eight public datasets, M2CG achieves state‑of‑the‑art results. The findings suggest that the reliance on recursion stems mainly from the lack of sufficient contextual information to fully characterize the remaining distance to the target alignment.
Review: The study demonstrates that a well‑designed one‑pass framework can match or surpass recursive methods, offering a more efficient solution for deformable registration.