Research map / Image generation / GANs
Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning
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 confusion-GAN, an unpaired virtual IHC staining method that combines a multi-branch confusion discriminator with a MIL-assisted patch-level pathology information extractor (PPIE) to keep pathological consistency.
Problem
Virtual immunohistochemistry (IHC) staining from H&E images usually needs hard-to-obtain paired (adjacent-slice) data, while unpaired methods often fail to preserve pathological (positive/negative) information.
Approach
Combines a CycleGAN-style unpaired translation framework with (1) a multi-branch discriminator whose confusion branch judges a generated image against a pool of real images, and (2) a PPIE that uses top-k MIL selection plus a double-spherical anomaly-detection loss, with a pathology loss enforcing consistency of positive/negative predictions between input H&E and generated IHC.
Key results
- On HCI, confusion-GAN reaches FID 77.3, LPIPS 0.570, SSIM 0.2300, beating CycleGAN (resnet) FID 90.9 and DCD FID 91.3.
- On BCI and MIST, the unpaired method achieves FID 49.2 and 39.8, lower than the supervised ASP baseline (54.3 and 41.2).
- Downstream ABMIL AUC for HCC classification rises from 55.6% using H&E images to 88.9% using generated GPC3 images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use confusion-GAN instead of CycleGAN (resnet) | virtual IHC staining · HCI | FID: 90.9 -> 77.3 | Helped |
| use confusion-GAN instead of ASP | virtual IHC staining · BCI | FID: 54.3 -> 49.2 | Helped |
| use confusion-GAN instead of ASP | virtual IHC staining · MIST | FID: 41.2 -> 39.8 | Helped |
| add PPIE to confusion-GAN (ablation) | confusion-GAN · HCI | FID: 105.4 -> 103.3 | Helped |
| add MBD to confusion-GAN (ablation) | confusion-GAN · HCI | FID: 105.4 -> 82.4 | Helped |
| add both PPIE and MBD (ablation) | confusion-GAN · HCI | FID: 105.4 -> 77.3 | Helped |
| use top-k selection with k=30 instead of without top-k selection (ablation) | PPIE patch classification with ABMIL · HCI patch classification | AUC: 0.732 -> 0.782 | Helped |
| use generated GPC3 images instead of H&E images | ABMIL · HCI downstream diagnosis | AUC: 55.6% -> 88.9% | Helped |
Surprising findings
- Pathologists labeled generated GPC3 images as real in roughly half of cases (accuracy 0.48-0.62 for fake/true), indicating near-chance discrimination.
- The unpaired method outperforms several fully supervised methods on FID and LPIPS on BCI and MIST, though its SSIM is lower on BCI.
Limitations
- Patch-level labels for BCI and MIST are derived via color analysis and label transfer rather than true WSI annotations.
- The top-k selection in PPIE degrades with large k, as negative patches contaminate the positive set.
- Evaluation of pathological consistency relies on three pathologists and a small test set of 20 H&E-IHC pairs.
Experimental setup
Datasets: HCI BCI MIST
Benchmarks: HCI BCI MIST
Metrics: FID LPIPS SSIM AUC FPR TPR G-mean pathologist accuracy
Baselines: CycleGAN UGATIT DCD Pix2pix Pix2pixHD PyramidP2P ASP
Builds on: CycleGAN ABMIL BBDM CLAM
Closest papers by meaning
- ODA-GAN: Orthogonal Decoupling Alignment GAN Assisted by Weakly-supervised Learning for Virtual Immunohistochemistry Staining (CVPR 2025 · similarity 0.92)
- OT-StainNet: Optimal Transport Driven Semantic Matching for Weakly Paired H&E-to-IHC Stain Transfer (AAAI 2025 · similarity 0.88)
- BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational Pathology (CVPR 2025 · similarity 0.83)
- Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation (CVPR 2025 · similarity 0.83)
- SheafStain: Sheaf-Theoretic Schr\"odinger Bridge for Spatially and Biologically Coherent Virtual Staining (NeurIPS 2026 · similarity 0.83)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.82)
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.