Research map / Image generation / Image restoration
A Plug-and-Play Image Registration Network
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 PIRATE, a plug-and-play deformable image registration method that uses a DnCNN denoiser trained on registration fields as an implicit regularizer, and PIRATE+, which fine-tunes that denoiser end-to-end via deep equilibrium models (JFB) for brain MRI registration.
Problem
Deep learning deformable image registration (DIR) networks output registration fields solely from a pre-trained CNN without enforcing data fidelity between the warped moving image and the fixed image at inference, and prior plug-and-play methods only regularize images, not registration fields.
Approach
Train a denoiser on registration fields and plug it as a regularizer into an iterative registration solver (PIRATE); then interpret the fixed-point iteration as an infinite-depth network and fine-tune the denoiser end-to-end with DEQ/JFB using an NCC + smoothness + Jacobian loss (PIRATE+).
Key results
- On OASIS, PIRATE+ reaches avg DSC 0.7952 with 0.0495% negative JD, versus NODEO's 0.7947 DSC and 0.0531% negative JD.
- On CANDI, PIRATE+ reaches avg DSC 0.7633 and 0.0660% negative JD, the best among compared methods.
- PIRATE (without DEQ) achieves DSC 0.7948 on OASIS with 32.01 s inference, competitive with iterative baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune the denoiser with DEQ (ablation) | PIRATE · OASIS-1 | Neg. JD %: 0.0567 (0.002) -> 0.0495 (0.001) | Helped |
| fine-tune the denoiser with DEQ (ablation) | PIRATE · CANDI | Neg. JD %: 0.0833 (0.002) -> 0.0660 (0.001) | Helped |
| add the denoiser regularizer (ablation) | penalty function plus smoothness regularizer · OASIS-1 | JD: 0.2091 (0.001) -> 0.0567 (0.002) | Helped |
| add the denoiser regularizer (ablation) | penalty function plus smoothness regularizer · CANDI | JD: 1.5847 (0.001) -> 0.0833 (0.002) | Helped |
| add the smoothness regularizer (ablation) | penalty function plus denoiser regularizer · OASIS-1 | JD: 0.1120 (0.001) -> 0.0495 (0.001) | Helped |
| add the smoothness regularizer (ablation) | penalty function plus denoiser regularizer · CANDI | JD: 1.1844 (0.001) -> 0.0660 (0.001) | Helped |
| use a dynamic instead of fixed step size (ablation) | PIRATE+ · OASIS-1 | Avg. DSC: 0.7902 (0.044) -> 0.7952 (0.029) | Helped |
| use a dynamic instead of fixed step size (ablation) | PIRATE+ · CANDI | Avg. DSC: 0.7556 (0.071) -> 0.7633 (0.027) | Helped |
| replace DnCNN with a TV denoiser (ablation) | PIRATE · OASIS-1 | Avg. DSC: 0.7948 (0.033) -> 0.7617 (0.054) | Hurt |
| replace DnCNN with a TV denoiser (ablation) | PIRATE · CANDI | Avg. DSC: 0.7624 (0.020) -> 0.7481 (0.049) | Hurt |
Surprising findings
- PIRATE+ outperforms NODEO even though NODEO generated the ground truth used to train the denoiser.
- The DEQ fine-tuned denoiser barely changes DSC relative to PIRATE (0.7952 vs 0.7948 on OASIS) but substantially reduces negative JD.
Limitations
- PIRATE+ training is much slower than traditional DL baselines (1821.9 s/epoch).
- Validation limited to brain MRI (OASIS-1, CANDI) and one lung CT dataset; other anatomies and cross-modal registration untested.
- Small training set of 100 image pairs per dataset.
Experimental setup
Datasets: OASIS-1 CANDI EMPIRE10
Benchmarks: OASIS-1 CANDI EMPIRE10
Metrics: Dice similarity coefficient ratio of negative Jacobian determinant inference time
Baselines: SyN VoxelMorph SYMNet GraDIRN Log-Demons NODEO
Builds on: plug-and-play priors SD-RED deep equilibrium models Jacobian-free DEQ NODEO
Closest papers by meaning
- Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.87)
- IIRP-Net: Iterative Inference Residual Pyramid Network for Enhanced Image Registration (CVPR 2024 · similarity 0.87)
- Correlation-aware Coarse-to-fine MLPs for Deformable Medical Image Registration (CVPR 2024 · similarity 0.87)
- CARL: A Framework for Equivariant Image Registration (CVPR 2025 · similarity 0.85)
- H-ViT: A Hierarchical Vision Transformer for Deformable Image Registration (CVPR 2024 · similarity 0.85)
- SACB-Net: Spatial-awareness Convolutions for Medical Image Registration (CVPR 2025 · similarity 0.85)
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.