Research map / Image generation / Editing and style transfer
OT-StainNet: Optimal Transport Driven Semantic Matching for Weakly Paired H&E-to-IHC Stain Transfer
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 OT-StainNet, a LoRA-fine-tuned one-step diffusion (SD-Turbo) generator trained adversarially with optimal-transport semantic matching and a semantic consistency loss for weakly paired H&E-to-IHC virtual staining.
Problem
Virtual IHC staining from H&E when adjacent-section H&E and IHC images are only weakly paired (misaligned at pixel level), so pixel-level losses introduce inaccurate pathological features.
Approach
Uses optimal transport to align H&E and adjacent IHC features semantically, then enforces that the pairwise feature correlation matrix of virtual IHC matches that of the OT-aligned real IHC features, on top of a pre-trained SD-Turbo generator adapted with LoRA.
Key results
- On MIST H&E→Ki67, OT-StainNet reports SSIM 0.158, PSNR 15.516, FID 36.286, outperforming the listed baselines on FID across all four stains.
- On IHC4BC H&E→PR, OT-StainNet reports SSIM 0.635 and FID 31.076, the best of the compared methods.
- Ablation on MIST shows removing OTSM and SCC raises FID from 36.286 to 47.331 (w/o SCC: 39.980) for Ki67.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add optimal transport-driven semantic matching (OTSM) (ablation) | OT-StainNet · MIST, H&E→Ki67 | SSIM: 0.138 -> 0.158 | Helped |
| Add optimal transport-driven semantic matching (OTSM) (ablation) | OT-StainNet · MIST, H&E→ER, PR, and HER2 | SSIM: 0.138 -> 0.148; 0.120 -> 0.142; 0.113 -> 0.118 | Helped |
| Add semantic consistency constraint (SCC) instead of standard L1 loss (ablation) | OT-StainNet · MIST, H&E→Ki67 | SSIM: 0.150 -> 0.158 | Helped |
| Add semantic consistency constraint (SCC) instead of standard L1 loss (ablation) | OT-StainNet · MIST, H&E→ER, PR, and HER2 | SSIM: 0.151 -> 0.148; 0.131 -> 0.142; 0.107 -> 0.118 | Helped |
Surprising findings
- Unsupervised CycleGAN-Turbo and CUT, and semi-supervised PRGAN, underperform in SSIM and PSNR, while fully supervised Pix2Pix-Turbo and PyramidP2P score higher on those metrics but poorly on FID.
Limitations
- Performance depends on dataset quality; contaminated training images where cell structures cannot be identified degrade results.
- Validated only on H&E-to-IHC for four stains (Ki67, ER, PR, HER2) on two datasets, with no special or fluorescence stain experiments.
Experimental setup
Datasets: MIST IHC4BC
Benchmarks: MIST IHC4BC
Metrics: SSIM PSNR FID Pearson correlation of positive area size
Baselines: Pix2Pix-Turbo CycleGAN-Turbo CUT PyramidP2P PRGAN ASP U-Frame
Builds on: img2img-turbo CycleGAN SD-Turbo CUT
Closest papers by meaning
- ODA-GAN: Orthogonal Decoupling Alignment GAN Assisted by Weakly-supervised Learning for Virtual Immunohistochemistry Staining (CVPR 2025 · similarity 0.89)
- Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning (CVPR 2024 · similarity 0.88)
- SheafStain: Sheaf-Theoretic Schr\"odinger Bridge for Spatially and Biologically Coherent Virtual Staining (NeurIPS 2026 · similarity 0.86)
- BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational Pathology (CVPR 2025 · similarity 0.85)
- Multi-Resolution Pathology-Language Pre-training Model with Text-Guided Visual Representation (CVPR 2025 · similarity 0.85)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · 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.