SAR2Net: Learning Spatially Anchored Representations for Retrieval-Guided Cross-Stain Alignment

Tianle Shen; Fang Yan; Xiaofan Zhang

cs.CVOct 19, 2026

SAR2Net: Learning Spatially Anchored Representations for Retrieval-Guided Cross-Stain Alignment

Authors: Tianle Shen, Fang Yan, Xiaofan Zhang

Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory

Abstract

Achieving spatial alignment between whole-slide images(WSIs) across stains is essential for integrating morphology and molecular information, yet it remains challenging due to extreme resolution, tissue fragmentation, and large nonlinear deformations. Conventional registration pipelines depend on global pre-alignment and spatial consistency, which often collapse under such distortions. We present SAR2Net, a framework that learns spatially anchored representations and reformulates cross-stain alignment as a region-level feature retrieval problem. Instead of estimating explicit transformations, SAR2Net learns pointwise representations encoding the relative spatial relationships to tissue landmarks. Given landmarks and arbitrary coordinates, it predicts spatially anchored features that serve as deformation-invariant descriptors of tissue topology. A twostage retrieval framework then establishes correspondences between slides, even when global alignment is infeasible. Experiments on biopsy-oriented HE-IHC datasets show that SAR2Net achieves robust region-level alignment under severe tissue distortions, outperforming previous methods. Code is available at https://github.com/Shentl/SAR2Net.

Figures & tables

Figure 1: Examples of biopsy slides. (A.1-A.2) HE and IHC slides of a core needle biopsy with tissue fractures and bending. (B.1-B.2) slides of a forceps biopsy with inter-block displacement.

Figure 2: Pipeline of SAR2Net. (a) Given source and target WSIs, we extract paired landmarks automatically as spatial anchors. As detailed in Sec. 3.3 , for each source patch, nearby anchors are selected via Spatially Coherent Anchor Grouping. Then for all points in both slides, SAR2Net computes spatially anchored representations that encode their relative geometry to the anchors. Afterwards, region correspondence is obtained through feature retrieval. If the highest similarity falls inside the source mask, the corresponding target points are assigned to the aligned mask. The retrieved regions are refined by a second round refinement. (b) As described in Sec. 3.2 , SAR2Net is trained via contrastive learning using purely synthetic point coordinates, combining random and local geometry–aware negative sampling.

Figure 3: Local geometry–aware negative sampling. (a) Example of a query (blue) near anchors (red), with LGN samples (black). (b) Sim. map w/o LGN. (c) Sim. map w/ our convex-hull based LGN, enabling the model to capture local geometric relations. In (b-c), the black circle highlights a local neighborhood around the reference point that is very close to the query, while the red circle marks a neighborhood around an anchor that is far from the query.

Figure 4: Two-Stage Refinement. The second-stage refinement is operated on each region and its neighborhood obtained from the first-stage retrieval. The $\circ$ indicates mismatched areas from the first stage that are successfully corrected after refinement.

Figure 5: Example of data annotation. The left and right slides are adjacent slides, where the right one exhibits severe distortion and has an additional control tissue. After segmenting the tissue foreground (green solid lines) and annotating the corresponding regions, four paired region masks were obtained.

Method	mIoU $\uparrow$	mDice $\uparrow$	aw-IoU $\uparrow$	aw-Dice $\uparrow$
VALIS [ 11 ]	0.635	0.705	0.647	0.716
RegWSI [ 28 ]	0.691	0.786	0.699	0.794
Ours (single)	0.891	0.938	0.893	0.939
Ours (second)	0.899	0.942	0.901	0.944

Table 1: Quantitative performance with mIoU, mDice, aw-IoU, and aw-Dice. Ours (single) means only the first round region retrieval and alignment is performed. Ours (second) means the second round refinement is performed.

	$\mathrm{SR}_{\mathrm{IoU}}^{0.75}\uparrow$	$\mathrm{SR}_{\mathrm{IoU}}^{0.85}\uparrow$	$\mathrm{SR}_{\mathrm{Dice}}^{0.75}\uparrow$	$\mathrm{SR}_{\mathrm{Dice}}^{0.85}\uparrow$
VALIS [ 11 ]	0.594	0.429	0.690	0.606
RegWSI [ 28 ]	0.545	0.324	0.740	0.574
Ours (single)	0.899	0.768	0.971	0.910
Ours (second)	0.939	0.812	0.974	0.946

Table 2: Robustness Analysis with success rate (SR) over IoU and Dice. Ours (single) means only perform the first round region retrieval and alignment. Ours (second) represents that the second round refinement is performed.

Figure 6: Visualization of the success rate curves. (a) Success rate curves over IoU thresholds. (b) Success rate curves over Dice thresholds. The AUC of the curves are reported on the figures.

w. pertu	w. LGN	w. sec	mIoU $\uparrow$	mDice $\uparrow$	aw-IoU $\uparrow$
✗	✓	✗	0.847	0.911	0.852
✓	✗	✗	0.888	0.936	0.890
✓	✓	✗	0.891	0.938	0.893
✓	✓	✓	0.899	0.942	0.901

Table 3: Ablation study on the presence of random perturbations (pertu), local geometry-aware negative sampling (LGN), and second-round refinement (sec).

Figure 7: Ablation study on random perturbations. (a) Similarity map generated by SAR2Net trained without random perturbation, (b) Similarity map by SAR2Net trained with random perturbation, with the query point shown in blue and anchor points in red.

Figure 8: We visualize three representative cases exhibiting fragmented tissues or strong deformations. In the first row, each slide contains multiple tissue fragments that undergo distinct spatial displacements. The second row shows a sample with a tissue piece undergoing large distortion. The third row presents intra-block misalignment, where an additional control tissue on the left (shown in the VALIS result) further distracts the registration. Our method achieves robust spatial correspondence across stains despite these challenges.

Figure 9: Success rate curves for ablation study. (a) Curves over IoU thresholds. (b) Curves over Dice thresholds. From left to right, each combination of ✓ and ✗ indicates whether random perturbations, Local Geometry-Aware Negative Sampling (LGN), and second-round refinement are enabled, respectively. The AUC of the curves are reported on the figures.

Medical Image Registration

Medical image registration aligns scans across time, patients, or imaging modalities to track anatomy and support diagnosis, treatment, and image-guided procedures.

See all 95 papers

INST-Align: Implicit Neural Alignment for Spatial Transcriptomics via Canonical Expression Fields
Apr 13, 2026Bonian Han, Cong Qi, Przemyslaw Musialski +1Batch Effect Correction Spatial Transcriptomics
New Jersey Institute of Technology, Newark, NJ, USA
Cluster-First Labelling: An Automated Pipeline for Segmentation and Morphological Clustering in Histology Whole Slide Images
Apr 10, 2026Muhammad Haseeb Ahmad, Sharmila Rajendran, Damion Young +1Whole-Slide Image Analysis Microscopy Image Analysis
Department of Engineering Science, University of Oxford, United Kingdom · Department of Physiology, Anatomy and Genetics, University of Oxford, United Kingdom · Medical Sciences Division, University of Oxford, United Kingdom
AINet: Anchor Instances Learning for Regional Heterogeneity in Whole Slide Image
Feb 21, 2026Tingting Zheng, Hongxun Yao, Kui Jiang +2Whole-Slide Image Analysis Multiple Instance Learning
Harbin Institute of Technology · Tsinghua University · Zhengzhou University

Apr 13, 2026J/K move · Enter open · S save

INST-Align: Implicit Neural Alignment for Spatial Transcriptomics via Canonical Expression Fields

Bonian Han, Cong Qi, Przemyslaw Musialski, Zhi Wei

New Jersey Institute of Technology, Newark, NJ, USA

Batch Effect Correction Spatial Transcriptomics Implicit Neural Representations

Open paper AskPDF arXiv

Spatial transcriptomics (ST) measures mRNA expression while preserving spatial organization, but multi-slice analysis faces two coupled difficulties: large non-rigid deformations across slices and inter-slice batch effects when alignment and integration are treated independently. We present INST-Align, an unsupervised pairwise framework that couples a coordinate-based deformation network with a shared Canonical Expression Field, an implicit neural representation mapping spatial coordinates to expression embeddings, for joint alignment and reconstruction. A two-phase training strategy first establishes a stable canonical embedding space and then jointly optimizes deformation and spatial-feature matching, enabling mutually constrained alignment and representation learning. Cross-slice parameter sharing of the canonical field regularizes ambiguous correspondences and absorbs batch variation. Across nine datasets, INST-Align achieves state-of-the-art mean OT Accuracy (0.702), NN Accuracy (0.719), and Chamfer distance, with Chamfer reductions of up to 94.9% on large-deformation sections relative to the strongest baseline. The framework also yields biologically meaningful spatial embeddings and coherent 3D tissue reconstruction. The code will be released after review phase.

Abstract

Medical Image Registration

Explore similar work

INST-Align: Implicit Neural Alignment for Spatial Transcriptomics via Canonical Expression Fields

Cluster-First Labelling: An Automated Pipeline for Segmentation and Morphological Clustering in Histology Whole Slide Images

AINet: Anchor Instances Learning for Regional Heterogeneity in Whole Slide Image