Any2Reg: Set-Based Groupwise Registration for Variable-Length, Variable-Contrast Cardiac MRI

MICCAI 2026 Early Acceptance (Top 9%)

Yi Zhang Yidong Zhao Tijmen Toxopeus Maša Božić-Iven Sebastian Weingärtner Qian Tao

Department of Imaging Physics, Delft University of Technology, the Netherlands

Code Paper (PDF) BibTeX
TL;DR Any2Reg (Any2Reg) is a set-based groupwise registration framework developed to make CMR registration more flexible and practical: it treats sequences as unordered sets, decouples the model from protocol-specific length and ordering, and shows strong cross-protocol zero-shot generalization on STONE, MOLLI, ASL, and Cine (L in [11, 60]).

Abstract

Cardiac MRI protocols present substantial heterogeneity in sequence length and contrast mechanisms, which challenges conventional groupwise registration models based on fixed channel stacking. We introduce a set-based groupwise registration formulation that decouples sequence order from feature extraction and constructs a canonical reference representation shared across frames. This design yields permutation-equivariant registration behavior and supports arbitrary-length input sequences without architecture changes. Across representative cardiac MRI datasets, our approach demonstrates improved registration quality and more reliable downstream mapping metrics in both qualitative and quantitative evaluations.

Method Overview

Method overview from paper
Set-based Any2Reg pipeline: shared feature encoding, correlation-guided aggregation, permutation-invariant reference construction, and equivariant deformation mapping. View source figure (PDF).

Qualitative Result

Qualitative result from paper on case 0138_4_0000
Head-to-head qualitative comparison on case 0138_4_0000. Any2Reg and Any2Reg-IO yield cleaner T1 maps, lower uncertainty, and higher fitting quality in regions with large motion. View source figure (PDF).
Slice-wise R2 analysis from paper
Slice-wise quantitative evidence on STONE (LV+Myo): Any2Reg-IO consistently improves mean R2 from base to apex, with stronger survival behavior across thresholds. View source figure (PDF).

Qualitative Result (anime!)

Method comparison animation on STONE
STONE (training, L=11). Multi-method comparison across temporal frames.
Method comparison animation on ACDC
ACDC (unseen, L=30). Multi-method comparison across temporal frames.

Availability

Code and pretrained checkpoints are available in this repository.

BibTeX

@inproceedings{zhang2026any2regnet,
  title     = {Set-Based Groupwise Registration for Variable-Length, Variable-Contrast Cardiac MRI},
  author    = {Zhang, Yi and Zhao, Yidong and Toxopeus, Tijmen and Bozic-Iven, Masa and Weingartner, Sebastian and Tao, Qian},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  year      = {2026}
}