Research map / Image generation / Image restoration
IIRP-Net: Iterative Inference Residual Pyramid Network for Enhanced 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 IIRP-Net, a streamlined 4-level pyramid registration network (RP-Net) whose residual flow estimators are reused iteratively coarse-to-fine at inference, with an NCC-based IterStop rule choosing iteration counts per image pair, for unsupervised deformable 3D medical image registration.
Problem
Deep learning-based deformable image registration networks are used in a single pass at inference and underexploit their trained registration capability, with pyramid networks showing limited accuracy on large deformations and fine structures.
Approach
Iteratively reapply the trained residual flow estimators of a pyramid network at each scale (feature-level internal iteration, coarse to fine) instead of repeating the whole network, with an adaptive IterStop criterion based on NCC change between successive outputs, adding no training cost.
Key results
- IIRP-Net reaches 73.3% Dice on FLARE and 65.8% on Mindboggle, versus 72.2% and 63.6% for the best prior baseline SDH-Net.
- RP-Net alone (410.1K params) gets 68.7% and 64.7% Dice; iterative inference adds about 4.6 and 1.2 Dice points on FLARE and Mindboggle at fixed Res×2 setting.
- IIRP-Net's training time is 31.9% and 31.3% of SDH-Net's, and inference time 19.4% and 14.6% of SDH-Net's, with unchanged training cost relative to RP-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add iterative inference to RP-Net (ablation) | RP-Net · FLARE | Dice (%): +4.66 | Helped |
| add iterative inference to RP-Net (ablation) | RP-Net · Mindboggle | Dice (%): +1.17 | Helped |
| add 2 ResBlocks to the residual flow estimators (ablation) | RP-Net · FLARE | Dice (%): 65.72 -> 68.68 | Helped |
| add 2 ResBlocks to the residual flow estimators (ablation) | RP-Net · Mindboggle | Dice (%): 63.24 -> 64.67 | Helped |
| increase residual flow estimator depth from 3 to 4 ResBlocks (ablation) | RP-Net · FLARE | Dice (%): 69.14 -> 69.04 | Hurt |
| increase residual flow estimator depth from 3 to 4 ResBlocks (ablation) | RP-Net · Mindboggle | Dice (%): 64.89 -> 64.97 | Helped |
| set IterStop threshold δ to 0.001 instead of 0.005 (ablation) | IIRP-Net · FLARE | Dice (%): 73.34 -> 73.47 | Helped |
| set IterStop threshold δ to 0.001 instead of 0.005 (ablation) | IIRP-Net · Mindboggle | Dice (%): 65.84 -> 65.74 | Hurt |
| use IIRP-Net instead of RP-Net | RP-Net · FLARE | Dice (%): 68.7 -> 73.3 | Helped |
| use IIRP-Net instead of RP-Net | RP-Net · Mindboggle | Dice (%): 64.7 -> 65.8 | Helped |
| use IIRP-Net instead of SDH-Net | SDH-Net · FLARE | Dice (%): 72.2 -> 73.3 | Helped |
| use IIRP-Net instead of SDH-Net | SDH-Net · Mindboggle | Dice (%): 63.6 -> 65.8 | Helped |
Surprising findings
- Fixed-count iteration of the whole network (RP-Net x t) or of all flow estimators degrades accuracy beyond a few iterations, so stopping is necessary.
- Under the same IterStop threshold, the number of iterations per scale differs substantially between datasets and scales.
Limitations
- IterStop relies on intensity-based NCC similarity, which may not align with anatomical structures.
- Excessive fixed iterations cause accuracy to drop sharply without a stopping criterion.
- Evaluated only on two 3D datasets (abdomen CT and brain MRI) with 2 ResBlocks chosen for efficiency rather than best accuracy.
Experimental setup
Datasets: FLARE Mindboggle-101
Benchmarks: FLARE Mindboggle-101
Metrics: Dice Jacobian determinant |Js|<=0 HD95 MSE training time inference time parameters
Baselines: VoxelMorph LKU-Net TransMorph ModeT DualPRNet++ NICE-Net PIViT Image2Grid RCN SDH-Net
Builds on: VoxelMorph LapIRN Dual-PRNet++ NCC loss Adam
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 0.84)
- Iterative Residual Refinement for Joint Optical Flow and Occlusion Estimation (CVPR 2019 · similarity 0.83)
- SACB-Net: Spatial-awareness Convolutions for Medical Image Registration (CVPR 2025 · similarity 0.83)
- CARL: A Framework for Equivariant Image Registration (CVPR 2025 · similarity 0.83)
- Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.82)
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.