Research map / 3D vision / Depth and stereo
SACB-Net: Spatial-awareness Convolutions for 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 SACB-Net, a 3D pyramid deformable image registration network whose spatial-awareness convolution block clusters features with K-Means and generates cluster-specific adaptive convolution kernels.
Problem
Learning-based deformable medical image registration relies on spatially shared convolution kernels, which poorly capture spatially varying information in non-local regions and hurt deformation field estimation, especially for large deformations.
Approach
A 3D Spatial-Awareness Convolution Block (SACB) clusters feature maps with unsupervised GPU K-Means on local-mean unfolded patches, then generates per-cluster convolution weights and biases from centroids via MLPs, and embeds it in a coarse-to-fine pyramid flow estimator with similarity matching.
Key results
- On IXI atlas-based registration, SACB-Net reaches 0.769 Dice vs 0.765 for LKU, the best baseline, with 1.11M parameters.
- On LPBA inter-subject registration, SACB-Net reaches 0.731 Dice vs 0.721 for ModeT.
- On abdominal CT, SACB-Net reaches 0.588 Dice vs 0.550 for ModeT and 0.444 for TransMorph.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SACB-Net instead of LKU | IXI atlas-based registration · IXI (30 ROIs) | Dice: 0.765˘0.129 -> 0.769˘0.127 | Helped |
| add SACB-Net instead of ModeT | LPBA inter-subject registration · LPBA (54 ROIs) | Dice: 0.721˘0.013 -> 0.731˘0.012 | Helped |
| add SACB-Net instead of ModeT | Abdomen CT inter-subject registration · Abdomen CT (13 ROIs) | Dice: 0.550˘0.054 -> 0.588˘0.049 | Helped |
| add SACB at scales 5–2 (ablation) | SACB-Net · IXI | Dice: 0.7643 -> 0.7683 | Helped |
| add SACB at scales 5–2 (ablation) | SACB-Net · LPBA | Dice: 0.7141 -> 0.7294 | Helped |
| add SACB at scales 5–2 (ablation) | SACB-Net · Abdomen CT | Dice: 0.5374 -> 0.5849 | Helped |
| use 7 clusters instead of 5 (ablation) | SACB-Net with SACB at scales 5–2 · LPBA | Dice: 0.7294 -> 0.7309 | Helped |
| use 9 clusters instead of 7 (ablation) | SACB-Net with SACB at scales 5–2 · LPBA | Dice: 0.7309 -> 0.7304 | Hurt |
| use 11 clusters instead of 7 (ablation) | SACB-Net with SACB at scales 5–2 · LPBA | Dice: 0.7309 -> 0.7300 | Hurt |
| use channel mean instead of spatial mean for cluster estimation (ablation) | SACB-Net with 5 clusters · LPBA | Dice: 0.7294 -> 0.7281 | Hurt |
| use spatial+channel mean instead of spatial mean for cluster estimation (ablation) | SACB-Net with 5 clusters · LPBA | Dice: 0.7294 -> 0.7285 | Hurt |
Limitations
- Abdomen CT |J|<0% of 3.930% is higher than several baselines.
- Ablation shows too many clusters (N=11) slightly decreases performance.
- Matching score is not applied at the finest scale due to memory and compute cost.
Experimental setup
Datasets: IXI LPBA Abdomen CT (Learn2Reg)
Benchmarks: IXI LPBA Abdomen CT
Metrics: Dice HD95 ASSD |J|<0% (Jacobian determinant folding)
Baselines: SyN VoxelMorph LKU TransMorph B-Spline-Diff Fourier-Net NCA-Morph LapIRN PRNet++ Im2Grid RDN ModeT
Builds on: VoxelMorph content-adaptive convolution Im2Grid ModeT LapIRN
Closest papers by meaning
- CARL: A Framework for Equivariant Image Registration (CVPR 2025 · similarity 0.86)
- Correlation-aware Coarse-to-fine MLPs for Deformable Medical Image Registration (CVPR 2024 · similarity 0.85)
- Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.85)
- A Plug-and-Play Image Registration Network (ICLR 2024 · similarity 0.85)
- H-ViT: A Hierarchical Vision Transformer for Deformable Image Registration (CVPR 2024 · similarity 0.84)
- 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.