Research map / Architectures and efficiency / Attention and transformers
Correlation-aware Coarse-to-fine MLPs for Deformable Medical Image Registration
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.
Proposes CorrMLP, the first MLP-based coarse-to-fine deformable medical image registration network, using correlation-aware multi-window MLP blocks at full resolution for brain MRI and cardiac MRI registration.
Problem
Transformers for deformable medical image registration are typically applied at downsampled resolutions, so they cannot capture fine-grained long-range dependence at full image resolution needed for precise dense correspondence; MLPs have not been well explored for registration and lack registration
Approach
A correlation-aware multi-window MLP (CMW-MLP) block that adds a 3D local correlation layer and parallel gMLP window branches (3, 5, 7 windows), used in a coarse-to-fine decoder that exploits both image-level correlations and step-level correlations between registration steps.
Key results
- CorrMLP reaches DSC 0.642 on Mindboggle and 0.661 on Buckner brain registration, above NICE-Trans (0.625 / 0.649).
- On ACDC cardiac registration CorrMLP achieves DSC 0.810 versus 0.799 for NICE-Trans, with NJD 0.389%.
- Ablation shows removing image-level or step-level correlation lowers DSC (e.g., Mindboggle 0.637 and 0.634 vs 0.642), and MLPMorph baseline reaches 0.604.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CorrMLP's correlation-aware coarse-to-fine architecture instead of a U-Net-like MLP architecture | MLPMorph · Mindboggle, Buckner, and ACDC | DSC: 0.604 -> 0.642; 0.632 -> 0.661; 0.780 -> 0.810 | Helped |
| Add image-level and step-level correlation information (ablation) | CorrMLP coarse-to-fine architecture · Mindboggle, Buckner, and ACDC | DSC: 0.628 -> 0.642; 0.650 -> 0.661; 0.800 -> 0.810 | Helped |
| Add step-level correlation information (ablation) | CorrMLP coarse-to-fine architecture with image-level correlation · Mindboggle, Buckner, and ACDC | DSC: 0.637 -> 0.642; 0.657 -> 0.661; 0.806 -> 0.810 | Helped |
| Add image-level correlation information (ablation) | CorrMLP coarse-to-fine architecture with step-level correlation · Mindboggle, Buckner, and ACDC | DSC: 0.634 -> 0.642; 0.655 -> 0.661; 0.805 -> 0.810 | Helped |
| Add a 3D correlation layer to MW-MLP blocks (ablation) | MW-MLP · Mindboggle, Buckner, and ACDC | DSC: 0.628 -> 0.642; 0.650 -> 0.661; 0.800 -> 0.810 | Helped |
| Add the 3×3×3 MLP branch (ablation) | CorrMLP CMW-MLP blocks · Mindboggle, Buckner, and ACDC | DSC: 0.639 -> 0.642; 0.657 -> 0.661; 0.808 -> 0.810 | Helped |
| Add the 5×5×5 MLP branch (ablation) | CorrMLP CMW-MLP blocks · Mindboggle, Buckner, and ACDC | DSC: 0.635 -> 0.642; 0.654 -> 0.661; 0.805 -> 0.810 | Helped |
| Add the 7×7×7 MLP branch (ablation) | CorrMLP CMW-MLP blocks · Mindboggle, Buckner, and ACDC | DSC: 0.637 -> 0.642; 0.655 -> 0.661; 0.806 -> 0.810 | Helped |
| Add a 9×9×9 MLP branch (ablation) | CorrMLP CMW-MLP blocks · Registration datasets | DSC | No clear effect |
Surprising findings
- Even without the correlation layer, the multi-window MLP block alone outperformed five existing MLP blocks (S2-MLP, sMLP, Hire-MLP, Swin-MLP, MAXIM's multi-axis gMLP).
- Adding an extra 9x9x9 MLP branch gave no further improvement.
Limitations
- MAXIM was too memory-heavy for 3D brain MRI and was evaluated only on cardiac data at smaller size.
- Evaluated on unsupervised NCC-based training only; multi-modal registration is proposed as future work but not tested.
Experimental setup
Datasets: ADNI ABIDE ADHD IXI Mindboggle Buckner ACDC
Benchmarks: Mindboggle Buckner ACDC
Metrics: Dice similarity coefficient negative Jacobian determinant percentage runtime
Baselines: SyN NiftyReg VoxelMorph Swin-VoxelMorph TransMorph TransMatch LapIRN ULAE-net Dual-PRNet++ NICE-Net NICE-Trans SDHNet
Builds on: gMLP NICE-Net MAXIM Dual-PRNet++ VoxelMorph
Closest papers by meaning
- A Plug-and-Play Image Registration Network (ICLR 2024 · similarity 0.87)
- Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.86)
- H-ViT: A Hierarchical Vision Transformer for Deformable Image Registration (CVPR 2024 · similarity 0.86)
- SACB-Net: Spatial-awareness Convolutions for Medical Image Registration (CVPR 2025 · similarity 0.85)
- CARL: A Framework for Equivariant Image Registration (CVPR 2025 · similarity 0.84)
- Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration (CVPR 2024 · 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.