Rethink-Reg: Reevaluating MRI Registration Methods

• The paper demonstrates that marginal gains in MRI registration arise from registration-specific modules rather than advanced computational backbones, achieving around a 1.5% Dice improvement.
• It employs a controlled experimental framework with consistent presets across methods (e.g., VXM, TM, Mam-TM) to disentangle low-level substitutions from registration-specific designs.
• The study advocates using comprehensive evaluation metrics including overlap and deformation quality, challenging the trend of attributing progress solely to trendy backbones.

Rethink-Reg denotes the reevaluation framework and code release associated with the study “Mamba? Catch The Hype Or Rethink What Really Helps for Image Registration,” which examines whether recent “advanced” computational blocks such as Transformers and Mamba actually improve deformable medical image registration under controlled comparisons (Jian et al., 2024). In the available arXiv record, the name appears through the repository rethink-reg, while the paper’s substantive contribution is a modular analysis of low-level backbone substitutions versus high-level registration-specific designs in mono-modal cross-sectional brain MRI registration (Jian et al., 2024). The study argues that registration gains are often misattributed when backbone changes are introduced together with coarse-to-fine pyramids, warping, correlation, and iterative refinement, and it reports that the latter, rather than trend-driven computational blocks, account for the measurable improvements (Jian et al., 2024).

1. Conceptual framing and research question

Rethink-Reg is organized around a specific methodological criticism: learning-based registration papers often import the latest computer vision backbone and present improved results without adequately disentangling those low-level substitutions from registration-specific architectural mechanisms (Jian et al., 2024). The targeted trend is the replacement of CNNs by attention-based Transformers, Mamba or selective state-space models, and large-kernel CNN variants, typically justified by their success in other vision tasks rather than by task-specific evidence in registration (Jian et al., 2024).

The central question is therefore twofold. First, do newer computational blocks materially improve deformable registration accuracy? Second, if improvements occur, do they originate instead from designs that explicitly encode registration structure, such as coarse-to-fine pyramids, warping of moving features, correlation volumes, and iterative refinement (Jian et al., 2024)? The paper’s answer is that advanced computational elements do not significantly improve registration accuracy under fair, controlled comparisons, whereas registration-specific modules provide modest but consistent gains, typically around a marginal $1.5\%$ Dice improvement over the baseline on several datasets (Jian et al., 2024).

This framing places Rethink-Reg less as a single novel registration operator than as an evaluation program. Its contribution is empirical and methodological: it argues that registration progress should be attributed by component-level isolation rather than by architectural fashion (Jian et al., 2024).

2. Controlled experimental framework

The study uses VoxelMorph (VXM) as the vanilla baseline, implemented as a U-Net-like network that concatenates source and target images at the input and predicts a dense deformation field $\phi$ through $F_\theta(I_s, I_t)=\phi$ (Jian et al., 2024). Downsampling and upsampling are implemented with stride-2 convolutions and transposed convolutions rather than max-pooling and trilinear interpolation, because the authors report that these choices improve performance (Jian et al., 2024). The final displacement field is predicted at half resolution, then upsampled and scaled to warp the full-resolution source image and label (Jian et al., 2024).

A defining feature of Rethink-Reg is its emphasis on comparison fairness. The paper states that all methods use the same data domain, the same training presets, the same random seed, the same dataloader behavior, and the same image pairs in each training iteration (Jian et al., 2024). All methods are trained for 100 epochs with learning rate $10^{-4}$, exponential decay $0.996$, random seed 2023, and loss weights $\gamma=\lambda=0.5$ (Jian et al., 2024). For testing, 200 random pairs are sampled from each dataset, and LPBA and Mindboggle are treated as zero-shot datasets with no training on them (Jian et al., 2024).

The task setting is mono-modal cross-sectional T1-weighted brain MRI deformable registration across five open-source datasets: OASIS, ADNI, IXI, LPBA, and Mindboggle (Jian et al., 2024). Reported splits are 330/84 for OASIS, 234/43 for ADNI, 301/115 for IXI, 0/40 for LPBA, and 0/100 for Mindboggle (Jian et al., 2024).

A practical consequence of this setup is that parameter count is not equalized across methods. VXM has 2.6M parameters, TM 46.6M, Mam-TM 39.0M, LKU 8.3M, and DWCPI 8.1M (Jian et al., 2024). This suggests that Rethink-Reg is not a strict parameter-matched benchmark; instead, it asks whether larger or trendier blocks win under otherwise fixed training and data conditions.

3. Architectural decomposition

The paper separates candidate improvements into low-level computational blocks and high-level registration-specific designs (Jian et al., 2024). This distinction is the core structural idea of Rethink-Reg.

Variant	Main modification	Role in study
VXM	CNN baseline	Vanilla reference
TM	Swin-Transformer-style encoder	Transformer substitution
Mam-VXM	Mamba blocks in VXM encoder	Mamba substitution
Mam-TM	Mamba blocks replacing TM encoder blocks	Mamba vs Transformer comparison
LKU	Large-kernel convolution blocks	Larger receptive field without attention
Dual	Two shared-weight encoders	Registration-specific feature separation
DWP	Dual + warping + pyramid	Coarse-to-fine motion estimation
DWCP	DWP + correlation	Explicit matching
DWCPI	DWCP + iteration	Iterative refinement

The low-level substitutions are designed to test whether better generic representation learning helps registration. TM replaces CNN encoder blocks with Transformer blocks following the TransMorph direction, Mam-VXM replaces CNN encoder blocks with Mamba blocks, Mam-TM replaces TM’s Swin blocks with Mamba blocks, and LKU replaces encoder blocks with large-kernel CNN blocks (Jian et al., 2024). These variants ask whether long-range modeling or enlarged receptive fields improve registration when isolated from other architectural changes.

The registration-specific branch introduces modules that more directly match the correspondence-estimation structure of registration (Jian et al., 2024). The dual-stream encoder processes target and source images separately with shared weights instead of concatenating them at input. Motion pyramid plus warping performs coarse-to-fine prediction at resolutions $[1/16, 1/8, 1/4]$, with finer moving features warped by coarser estimated flow (Jian et al., 2024). Correlation explicitly computes feature similarity, with global correlation at $1/16$ resolution and local correlation with radii $[3,2,1]$ at $[1/8,1/4,1/2]$ (Jian et al., 2024). Iteration applies repeated refinement only at the last two levels, using 2 iterations at resolutions $\phi$0 (Jian et al., 2024).

This decomposition is central to the paper’s argument. It does not deny that sophisticated blocks may be useful in other domains; rather, it claims that in brain MRI registration the dominant gains come from mechanisms that explicitly encode matching, multi-scale motion, and refinement (Jian et al., 2024).

4. Objective function, correlation, and evaluation metrics

The training objective is reported as

$\phi$1

where $\phi$2 and $\phi$3 are target and source images, $\phi$4 and $\phi$5 are their labels, $\phi$6 is the dense deformation field, and both $\phi$7 and $\phi$8 are set to $\phi$9 (Jian et al., 2024). The paper names the three components as image similarity, segmentation Dice loss, and deformation smoothness regularization, but does not provide explicit formulas for the individual terms in the available text (Jian et al., 2024).

For correlation-based matching, after flattening spatial dimensions the target and source features are

$F_\theta(I_s, I_t)=\phi$0

and correlation is computed by

$F_\theta(I_s, I_t)=\phi$1

with $F_\theta(I_s, I_t)=\phi$2 (Jian et al., 2024). This provides an explicit matching signal rather than relying on the backbone to learn correspondences implicitly.

For pyramid outputs, similarity and regularization losses are computed at each level and weighted by resolution; the paper gives as an example that at level $F_\theta(I_s, I_t)=\phi$3, with resolution $F_\theta(I_s, I_t)=\phi$4, the loss term is $F_\theta(I_s, I_t)=\phi$5 (Jian et al., 2024). This indicates that coarser scales contribute to training without dominating the objective.

Evaluation goes beyond conventional registration accuracy. The study reports Dice score (DSC) and HD90, but also SDlogJ and NDV, where SDlogJ measures the standard deviation of the logarithm of the Jacobian determinant and NDV measures the percentage of non-diffeomorphic voxels within the brain area (Jian et al., 2024). This broader metric set is one of the paper’s methodological claims: accuracy differences between strong baselines can be small, while deformation plausibility may differ substantially (Jian et al., 2024).

5. Main empirical findings

The principal result is negative with respect to architectural hype. Transformers and Mamba do not consistently outperform the CNN baseline, despite much larger parameter counts in some cases (Jian et al., 2024). TM, with 46.6M parameters, generally underperforms VXM; Mam-TM also underperforms VXM; Mam-VXM occasionally gives tiny gains but not meaningful overall improvements; and LKU sometimes matches or slightly exceeds VXM on Dice while often producing worse deformation-regularity metrics (Jian et al., 2024).

By contrast, the progressive addition of registration-specific modules yields a consistent improvement pattern. Dual-stream encoding alone does not help much and can slightly hurt Dice, but adding pyramid plus warping improves performance, adding correlation improves it further, and adding iteration produces the strongest overall model, DWCPI (Jian et al., 2024).

Representative Dice results show this pattern clearly. On OASIS, VXM achieves $F_\theta(I_s, I_t)=\phi$6 and DWCPI $F_\theta(I_s, I_t)=\phi$7; on ADNI, $F_\theta(I_s, I_t)=\phi$8 versus $F_\theta(I_s, I_t)=\phi$9; on IXI, $10^{-4}$0 versus $10^{-4}$1; and on Mindboggle, $10^{-4}$2 versus $10^{-4}$3 (Jian et al., 2024). These correspond to the paper’s characterization of a marginal $10^{-4}$4 improvement over baseline on several datasets (Jian et al., 2024).

LPBA is the main exception, and also the strongest zero-shot case. There, VXM reports $10^{-4}$5, while DWCPI reaches $10^{-4}$6, with HD90 improving from $10^{-4}$7 to $10^{-4}$8, and NDV dropping from $10^{-4}$9 to $0.996$0 (Jian et al., 2024). This suggests that coarse-to-fine warping, correlation, and iteration may matter more under harder out-of-distribution conditions.

The deformation-quality metrics are central to the paper’s broader claim. On Mindboggle, VXM reports NDV $0.996$1 and DWCPI $0.996$2; on OASIS, NDV falls from $0.996$3 to $0.996$4 (Jian et al., 2024). This supports the argument that methods with similar Dice can differ meaningfully in topology preservation and regularity, and that Rethink-Reg’s emphasis on SDlogJ and NDV is not merely diagnostic but part of the substantive conclusion (Jian et al., 2024).

6. Methodological significance

Rethink-Reg’s main significance is methodological rather than algorithmic. The study argues that the field should not interpret architectural novelty as evidence of causal improvement unless low-level and high-level components are disentangled under fixed conditions (Jian et al., 2024). This is why the paper emphasizes same-domain training, same random seed, same dataloader behavior, same image-pair sampling, and a modular progression from VXM to Dual, DWP, DWCP, and DWCPI (Jian et al., 2024).

A second methodological contribution is the insistence that registration should be evaluated with metrics that capture deformation quality, not only overlap. The paper explicitly advocates “novel evaluation metrics that go beyond conventional registration accuracy,” and its results show cases where LKU has Dice similar to VXM but worse SDlogJ and NDV (Jian et al., 2024). This implies that overlap alone can obscure clinically relevant differences in deformation plausibility.

A third implication is practical. The study recommends simple VoxelMorph for many brain image registration problems and DWCPI when marginally higher accuracy is needed (Jian et al., 2024). This suggests that the most reliable route to improvement is not backbone escalation, but explicit modeling of matching structure through pyramids, warping, correlation, and refinement.

The paper also states several limitations. Its conclusions are drawn for brain MRI, mono-modal, cross-sectional, deformable registration; they are not claimed to generalize across all organs and modalities (Jian et al., 2024). The study does not report statistical significance tests, does not fully specify some loss formulas in the available text, and does not combine registration-specific designs with all advanced backbones, so it stops short of proving that Transformers or Mamba would remain unhelpful under every possible high-level architecture (Jian et al., 2024).

7. Terminology and relation to other “REG” usages

In the arXiv material considered here, the explicit string “rethink-reg” is directly associated with the code release for the image-registration study (Jian et al., 2024). This suggests that Rethink-Reg is best understood as the shorthand for that registration-focused reevaluation framework rather than as a generic algorithmic family name.

The term should be distinguished from several unrelated “REG” labels in other fields. “REG” can denote “Rectified Gradient Guidance” for conditional diffusion models (Gao et al., 31 Jan 2025), “Refined Generalized Focal Loss” for road asset detection (Panboonyuen, 2024), “Regularized End-point Gradient” thermodynamic integration (Kapil, 12 Feb 2026), the registry-driven enterprise architecture REGAL (Agrawal, 3 Mar 2026), or the cortical rigid-registration method NC-Reg (Vati et al., 26 Jan 2026). None of these is the subject of the image-registration reevaluation embodied by Rethink-Reg.

Within medical image registration specifically, Rethink-Reg is most precisely characterized as an evidence-based argument that progress should be attributed to registration-specific inductive bias and rigorous evaluation, rather than to the uncritical adoption of whatever backbone is currently fashionable (Jian et al., 2024).