Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Adaptation via Style and Structure Guidance for Histological 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 SGTTA, a test-time adaptation method for histological image registration using style guidance (AdaIN with feature statistics), HOG-based structure guidance, and a historical-model smoothing strategy, applied to U-Net, Pyramid, and Cascade backbones on ANHIR.
Problem
Learning-based histological image registration generalizes poorly to test images due to appearance variances (multi-stain) and local structure changes (repetitive textures, missing sections).
Approach
Test-time adaptation for registration that aligns encoder feature statistics between training and test styles via AdaIN, adds an HOG-based structural similarity constraint during per-pair adaptation, and smooths the initialization with a moving average of historical models.
Key results
- HistRegU AArTRE improves from 0.0254 to 0.0196 with SGTTA (AArTRE) and MArTRE from 0.0084 to 0.0051.
- HistRegC AMrTRE improves from 0.0214 to 0.0155 and HistRegP from 0.0230 to 0.0149 with SGTTA (per text).
- Ablation: style guidance improves AArTRE 0.0254 to 0.0231; full combination of style, structure and continuity reaches 0.0196.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SGTTA to HistRegU | HistRegU · ANHIR | AMrTRE: 0.0257 -> 0.0189 | Helped |
| Apply SGTTA to HistRegC | HistRegC · ANHIR | AMrTRE: 0.0214 -> 0.0155 | Helped |
| Apply SGTTA to HistRegP | HistRegP · ANHIR | AMrTRE: 0.0223 -> 0.0149 | Helped |
| Add style guidance (ablation) | HistRegU test-time adaptation · ANHIR | AArTRE: 0.0254 -> 0.0231 | Helped |
| Add structure guidance to style guidance (ablation) | HistRegU test-time adaptation · ANHIR | AArTRE: 0.0231 -> 0.0209 | Helped |
| Add model continuity to style and structure guidance (ablation) | HistRegU test-time adaptation · ANHIR | AArTRE: 0.0209 -> 0.0196 | Helped |
| Add encoded feature position 2 to style guidance (ablation) | HistRegU style guidance · ANHIR | AArTRE: 0.0207 -> 0.0203 | Helped |
| Add encoded feature position 3 to style guidance (ablation) | HistRegU style guidance · ANHIR | AArTRE: 0.0203 -> 0.0198 | Helped |
| Add encoded feature position 4 to style guidance (ablation) | HistRegU style guidance · ANHIR | AArTRE: 0.0198 -> 0.0196 | Helped |
| Add HOG scale 1/2 to structure guidance (ablation) | HistRegU structure guidance · ANHIR | AArTRE: 0.0219 -> 0.0206 | Helped |
| Add HOG scale 1/4 to structure guidance (ablation) | HistRegU structure guidance · ANHIR | AArTRE: 0.0206 -> 0.0199 | Helped |
| Add HOG scale 1/8 to structure guidance (ablation) | HistRegU structure guidance · ANHIR | AArTRE: 0.0199 -> 0.0196 | Helped |
Surprising findings
- Test-time adaptation adds only about 0.1 to 0.7 minutes per pair, yet is faster than most traditional methods while more accurate.
Limitations
- Evaluated only on one histological dataset (ANHIR subset with landmarks).
- Inference time is longer than pure learning-based one-step registration.
- Hyperparameters alpha, w, k set empirically.
Experimental setup
Datasets: ANHIR
Benchmarks: ANHIR
Metrics: rTRE (median-median, average-median, median-average, average robustness inference time
Baselines: bUnwarpJ RVSS NiftyReg Elastix ANTs DROP HistRegU HistRegP HistRegC SSMSR
Builds on: VoxelMorph DeepHistReg RDN RCN SSMSR Tent
Closest papers by meaning
- Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation (AAAI 2025 · similarity 0.84)
- Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation (CVPR 2025 · similarity 0.84)
- Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation (NeurIPS 2026 · similarity 0.84)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.84)
- Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.84)
- Quilt-LLaVA: Visual Instruction Tuning by Extracting Localized Narratives from Open-Source Histopathology Videos (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.