Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Learning in Medical Image Registration: Magic or Mirage?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Empirically compares classical optimization-based and deep learning image registration on brain MRI, linking classical Dice to intensity-label mutual information and showing DLIR gains from label supervision do not transfer across datasets.
Problem
Unclear when learning-based deformable image registration outperforms classical optimization-based methods, in accuracy and robustness to domain shift.
Approach
Shows classical registration Dice correlates strongly with intensity-label mutual information, argues the Jacobian/architecture of DLIR cannot improve label matching beyond this under image-matching objectives, finds label-supervised DLIR beats classical methods in-domain but fails to generalize across datasets, and proposes a recipe for choosing paradigms.
Key results
- Dice score of ANTs across four brain datasets correlates strongly with intensity-label mutual information (Pearson r=0.886 linear, r=0.933 logarithmic).
- Classical methods (e.g., Greedy, ANTs) are significantly better than unsupervised DLIR methods on OASIS (p<0.01 in most cells); top classical Greedy and top DLIR TransMorph are not significantly different (p=0.161).
- With label supervision, state-of-the-art DLIR methods significantly outperform classical methods in-domain on OASIS, but this gain does not transfer to CUMC12, LPBA40, MGH10, or IBSR18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Relate image-label mutual information to classical registration performance | ANTs · OASIS, LPBA40, MGH10, and IBSR18 | Pearson correlation with Dice score (linear): r = 0.886 | Helped |
| Relate image-label mutual information to classical registration performance | ANTs · OASIS, LPBA40, MGH10, and IBSR18 | Pearson correlation with Dice score (logarithmic): r = 0.933 | Helped |
| Use unsupervised DLIR instead of classical registration | OASIS · OASIS, val and trainval | Dice score: p=0.161 | Mixed |
| Add label-matching supervision during training | DLIR methods · OASIS, val | Dice score | Helped |
| Add label-matching supervision during training | DLIR methods · OASIS, trainval | Dice score | Helped |
| Train on OASIS and evaluate on other brain datasets | DLIR methods · CUMC12, LPBA40, MGH10, and IBSR18 | Mean Dice score | Hurt |
| Evaluate classical methods across datasets rather than only on OASIS | FireANTs, Greedy, ANTs, and NiftyReg · CUMC12, LPBA40, MGH10, and IBSR18 | Mean Dice score | Helped |
| Add label-matching supervision to OASIS-trained DLIR models | DLIR methods · Four T1-brain datasets | Dice score | No clear effect |
| Use the supervised SymNet variant instead of its unsupervised counterpart | SymNet · OASIS | Dice score | Helped |
| Use SynthMorph's domain-agnostic training paradigm | SynthMorph · CUMC12, LPBA40, MGH10, and IBSR18 | Mean Dice score | Helped |
Surprising findings
- Published baseline Dice scores were reported 2 to 10 points lower than the authors' re-evaluation, which they attribute to instrumentation bias.
- Label-supervised DLIR gains on OASIS do not transfer to other datasets, even compared to their own unsupervised variants.
Limitations
- Does not consider hybrid methods or correlation-volume or sparse-correspondence representations.
- Evaluation restricted to neuroimaging brain MRI; multimodal registration not studied.
- Results on OASIS use only 364 training and 50 validation images.
Experimental setup
Datasets: OASIS CUMC12 LPBA40 MGH10 IBSR18
Benchmarks: OASIS CUMC12 LPBA40 MGH10 IBSR18
Metrics: Dice score mutual information Pearson correlation paired two-sided t-test p-values
Baselines: SynthMorph LapIRN SymNet LKU-Net TransMorph NiftyReg Symmetric Log Demons Greedy FireANTs
Builds on: VoxelMorph ANTs SynthMorph
Closest papers by meaning
- A Plug-and-Play Image Registration Network (ICLR 2024 · similarity 0.87)
- Correlation-aware Coarse-to-fine MLPs for Deformable Medical Image Registration (CVPR 2024 · similarity 0.86)
- CARL: A Framework for Equivariant Image Registration (CVPR 2025 · similarity 0.86)
- SACB-Net: Spatial-awareness Convolutions for Medical Image Registration (CVPR 2025 · similarity 0.85)
- Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration (CVPR 2024 · similarity 0.85)
- Uncertainty Estimation for Pretrained Medical Image Registration Models via Transformation Equivariance (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.