Source-linked AI summary
StainBridge: Stain-Aware Pairwise Registration of Serial Renal Biopsy Whole-Slide Images Across Structural and Immunohistochemical Stains
Ellen Wei, Bohang Jiang, Yanfan Zhu, Daniel Reisenbüchler, Kenji Ikemura, Steven Salvatore, Surya Seshan, Thangamani Muthukumar, Mert R. Sabuncu, Yihe Yang, Ruining Deng
TL;DR
Serial renal biopsy WSIs are difficult to register across structural and IHC stains because marker expression can be sparse or absent, limiting shared visual cues needed for future 3D analysis. StainBridge combines stain-aware preprocessing with affine and nonrigid registration and systematically evaluates their interaction. DeeperHistReg provides the broadest coverage and strongest pooled landmark accuracy, while preprocessing benefits depend on the backend, stain pairing, and metric.
Problem
Sparse or absent IHC expression and varied structural stains leave limited shared appearance cues for registering serial renal biopsy WSIs, a setting not directly represented by existing benchmarks.
Method
StainBridge evaluates stain deconvolution, intensity normalization, and tissue-mask injection alongside XFeat affine initialization and four nonrigid registration backends.
Results
DeeperHistReg provided the broadest coverage and strongest pooled landmark accuracy, while preprocessing improved XFeat-initialized registration but did not consistently benefit DeeperHistReg.
Takeaways & Limitations
Preprocessing and backend selection should be matched to stain pairing and evaluation metric when registering serial renal biopsy images.
Takeaways & Limitations
The study is limited by its 23-case single-center cohort, relatively small structural–IHC subset, fallback grayscale processing for several markers, and evaluation of pairwise 2D rather than composed 3D registration.
Abstract
from arXiv · showhide
Three-dimensional (3D) reconstruction of histopathology tissue requires accurate pairwise registration of serial whole-slide images (WSIs). Cross-stain benchmarks have advanced registration of differently stained histology, including structural-to-immunohistochemistry (IHC) pairs, but serial renal biopsy stacks remain difficult: they interleave several structural stains with diverse IHC markers whose expression can be sparse or absent, leaving few shared features to match. We present StainBridge, a stain-aware framework for registering serial renal biopsy WSIs across structural and IHC stains. StainBridge couples three preprocessing components, stain deconvolution, intensity normalization, and tissue-mask injection, with XFeat-based affine initialization and four nonrigid backends (VoxelMorph, ConvexAdam, FireANTs, and DeeperHistReg). We evaluate it on 23 cases comprising 338 WSIs, four structural stains, and ten IHC markers, with functional tissue units annotated on consecutive sections to give 1,468 landmark correspondences across 272 image pairs, and report tissue-mask Dice, functional-unit centroid error in micrometers, and tissue-restricted structural similarity. Nonrigid refinement improves on the affine initialization for three of four backends, VoxelMorph being the exception. DeeperHistReg, which computes its own initialization rather than relying on XFeat, gives the best pooled landmark accuracy and registers the most pairs, including every attempted structural-IHC pair. Preprocessing improves landmark accuracy for ConvexAdam and FireANTs in every stain-pairing category, and FireANTs shows both the largest single preprocessing gain on structural-IHC pairs and the best pooled tissue overlap. These results offer practical guidance for cross-stain registration and a foundation for integrated 3D analysis of renal tissue architecture and molecular expression.
1. INTRODUCTION
Serial renal biopsy WSI registration enables future 3D association of tissue units and marker expression, but cross-stain renal stacks pose distinctive correspondence challenges. StainBridge addresses this gap by systematically evaluating stain-aware preprocessing and registration backends.
- Accurate pairwise alignment enables future 3D association of glomeruli and other functional tissue units with molecular marker expression.The study evaluates pairwise registration rather than composing a full 3D reconstruction.
- Sparse or absent IHC expression, multiple small tissue cores, and stain-dependent appearance create a distinct renal cross-stain registration problem not directly represented by existing benchmarks.Serial registration must also accommodate displacement, deformation, missing tissue, and scanning effects.
- StainBridge combines deconvolution, intensity normalization, and tissue-mask injection with XFeat affine initialization and four nonrigid backends.The framework evaluates preprocessing components individually and in combination rather than assuming a universal configuration.
2. METHOD
StainBridge preprocesses paired structural and IHC WSIs before global and local registration. It compares independently configurable stain-aware operations, shared XFeat initialization for three backends, and independent end-to-end alignment in DeeperHistReg.
- Cross-Stain Preprocessing: Three independently configurable operations—stain deconvolution, tissue-mask injection, and intensity normalization—reduce stain-dependent appearance differences and suppress background or artifact effects.Eight factorial configurations, including raw input, allow component effects to vary by stain pairing and method.
- Registration Pipelines: XFeat estimates global translation, rotation, scale, and shear before ConvexAdam, FireANTs, or VoxelMorph estimate local deformation from the affine-aligned images.These three backends share the selected category-specific affine initialization.
- Nonrigid Backends: ConvexAdam uses MIND-SSC, FireANTs uses multiscale local normalized cross-correlation, and VoxelMorph uses an untuned pretrained 3D brain-MRI model adapted to 2D inputs.Each backend represents a different registration family and produces a 2D deformation from its processed inputs.
- Registration Pipelines: DeeperHistReg is evaluated independently because it performs its own feature-based initialization and nonrigid refinement rather than accepting the XFeat transformation.Its preprocessed condition receives externally deconvolved images before internal grayscale and contrast normalization.
3. DATA & EXPERIMENTS
The evaluation uses a multi-stain renal biopsy cohort with glomerular landmarks and complementary overlap, distance, similarity, and runtime measurements. Registration transformations are applied consistently to images, tissue masks, and landmarks.
- Dataset: The dataset contains 338 WSIs from 23 renal biopsies spanning four structural stains and ten IHC markers, with 1,468 glomerular correspondences across 272 consecutive pairs.Slides were digitized at 40× magnification, and pairs lacking persistent visible glomeruli were excluded from landmark evaluation.
- Preprocessing: Deconvolution uses stain-specific or generic vectors when available, while PAMS, C5b-9, C4d, FoxP3, and CMV use grayscale conversion as a fallback.Generic literature-derived vectors were used because stain-specific controls were unavailable.
- Runtime: Approximate per-pair runtime is reported separately from preprocessing, with timing measured around each registration call and field transfer forcing pending GPU computation to complete.The runtime table therefore concerns registration calls rather than preprocessing cost.
- Registration Evaluation: Each consecutive pair treats the later section as moving and the earlier section as fixed, while estimated fields transform the moving image, tissue mask, and landmarks into the fixed frame.This direction supports consistent composition across serial stacks.
- Evaluation: Registration quality is measured using functional-unit centroid distances in micrometres, tissue-mask Dice, landmark-mask IoU, and tissue-restricted structural similarity.Glomeruli are polygon-annotated, matched by name, and evaluated only when present on both slides.
4. RESULTS
Across affine and nonrigid registration, preprocessing generally improved XFeat-initialized pipelines, but benefits depended on stain pairing, metric, and backend. DeeperHistReg offered broader coverage and strong landmark accuracy, while comparisons were limited by unmatched initialization and cohort scope.
- Affine registration: 145.1 to 118.9 µm: XFeat preprocessing reduced pooled landmark-weighted centroid distance, while Dice increased from 0.679 to 0.686.The highest pooled Dice was 0.699 with deconvolution plus tissue-mask injection, but that configuration did not minimize centroid distance.
- Nonrigid registration: Nonrigid refinement improved on pooled affine centroid distance for ConvexAdam, FireANTs, and DeeperHistReg, but VoxelMorph remained less accurate than affine-only registration.This pattern shows that refinement gains were backend-dependent rather than universal.
- Registration across stain-pairing categories: Preprocessing choices were stain-pairing-specific: structural–structural pairs benefited from deconvolution with mask injection, structural–IHC pairs from normalization alone, and IHC–IHC pairs from deconvolution.For structural–structural pairs, centroid distance fell from 136.8 to 99.4 µm; structural–IHC normalization achieved Dice 0.664 and centroid distance 92.4 µm; IHC–IHC deconvolution reduced distance from 161.5 to 122.4 µm.
- Registration across stain-pairing categories: Structural–IHC pairs had fragile XFeat initialization, with XFeat-dependent backends registering 13 of 22 pairs versus all 21 available pairs for DeeperHistReg.This category also showed the largest preprocessing benefit, a 52.7% centroid-distance reduction.
- Nonrigid registration: DeeperHistReg provided broader coverage and stronger landmark accuracy in key stain-pairing categories, while XFeat-initialized backends showed complementary performance and preprocessing tradeoffs.The comparison is not strictly matched because DeeperHistReg used independent initialization and excluded one case after repeated out-of-memory failures.
5. NEW OR BREAKTHROUGH WORK TO BE PRESENTED
StainBridge systematically evaluates pairwise registration of serial renal biopsy WSIs across structural and IHC stains. Its main contribution is characterizing how stain-aware preprocessing interacts with affine and nonrigid methods, showing that preprocessing usually helps XFeat-initialized pipelines while independent initialization can improve coverage when shared appearance cues are weak.
- Contribution: StainBridge systematically evaluates pairwise registration across four structural stains and ten IHC markers in serial renal biopsy whole-slide images.The study focuses on renal biopsies with multiple small tissue cores, sparse or absent marker expression, and glomerulus-based evaluation.
- Contribution: The study characterizes how stain-aware preprocessing interacts with affine and nonrigid registration across structural–structural, structural–IHC, and IHC–IHC pairs.Its contribution is a controlled evaluation rather than a new registration algorithm.
6. CONCLUSION
StainBridge provides practical guidance for cross-stain renal registration, with DeeperHistReg offering the broadest coverage and strongest overall landmark accuracy. The findings support future validation and patient-level three-dimensional reconstruction.
- DeeperHistReg provided the broadest coverage and strongest overall landmark accuracy, while XFeat-initialized methods showed method-dependent tradeoffs.
- Cross-stain preprocessing improved pooled affine registration and reduced centroid distance for all three XFeat-initialized nonrigid backends, but benefits varied by stain pairing and did not consistently help DeeperHistReg.
- Structural–structural pairs achieved the strongest tissue overlap, structural–IHC pairs showed the greatest initialization fragility and largest preprocessing benefit, and IHC–IHC pairs retained lower overlap despite higher appearance similarity.
- These findings provide practical guidance for cross-stain renal registration and support future validation and patient-level three-dimensional reconstruction.