Research map / Architectures and efficiency / Attention and transformers

SheafStain: Sheaf-Theoretic Schr\"odinger Bridge for Spatially and Biologically Coherent Virtual Staining

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: Efficient image transformers

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 SheafStain, a Schrödinger Bridge virtual staining generator conditioned on Prov-GigaPath spatial and CLS tokens, with sheaf-style pixel and cocycle losses to reduce stitching artifacts in H&E-to-IHC translation.

Problem

Patch-wise virtual staining of H&E to IHC produces stitching artifacts and inconsistencies across patch boundaries when reassembling whole-slide-scale images, and VFM embeddings of overlapping patches are inconsistent due to context contamination.

Approach

Treats pathology VFM tokens as sheaf sections, injecting a per-position spatial map and neighborhood CLS token as conditioning in UNSB, and adds pixel sheaf and cocycle losses on overlaps plus adaptive extra reference patches at inference.

Key results

• On BCI HER2, SheafStain reaches FID 36.36 and KID 4.21 versus 67.63 and 29.30 for the strongest prior method, UNIStainNet.
• On HER2 Low/High classification on BCI with a ResNet-50 trained on translated images, SheafStain reaches 0.766 accuracy and 0.794 AUROC, exceeding UNIStainNet by 0.042 accuracy and 0.057 AUROC.
• Ablation shows spatial conditioning gives the largest single-step gain, reducing BCI FID from 227.79 (UNSB) to 55.25.

Surprising findings

• UNSB baseline, the starting point of SheafStain, performs worst or near-worst on several BCI metrics, with FID 227.79 and HER2 classification accuracy 0.524.
• Tiling Score plateaus after the pixel sheaf loss, suggesting pairwise supervision saturates boundary consistency at 1024x1024.

Limitations

• Validation restricted to BCI and MIST; no external clinical or WSI-level evaluation.
• Cross-organ generalization is empirically unverified.
• Inference is computationally expensive due to VFM forward pass.

Experimental setup

Datasets: BCI MIST

Benchmarks: BCI MIST

Metrics: FID KID LPIPS DISTS SSIM PSNR Tiling Score DAB-r DAB KL DAB JSD

Baselines: pix2pix CycleGAN PSPStain D-VST UNIStainNet UNSB

Builds on: Unpaired Neural Schrödinger Bridge Prov-GigaPath Neural Sheaf Diffusion CycleGAN pix2pix

Closest papers by meaning

• Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.86)
• OT-StainNet: Optimal Transport Driven Semantic Matching for Weakly Paired H&E-to-IHC Stain Transfer (AAAI 2025 · similarity 0.86)
• HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis (NeurIPS 2024 · similarity 0.85)
• PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.84)
• Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.84)
• ODA-GAN: Orthogonal Decoupling Alignment GAN Assisted by Weakly-supervised Learning for Virtual Immunohistochemistry Staining (CVPR 2025 · similarity 0.84)