Research map / Label-efficient and robust learning / Recognition and long-tail
ODA-GAN: Orthogonal Decoupling Alignment GAN Assisted by Weakly-supervised Learning for Virtual Immunohistochemistry Staining
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 ODA-GAN, an unpaired virtual IHC staining GAN that uses UNI-based weakly-supervised segmentation, orthogonal MLP feature decoupling, dual-stream PatchNCE, and multi-layer MMD domain alignment.
Problem
Virtual immunohistochemistry (IHC) staining from H&E images suffers from staining unreality and unreliability: paired data is hard to align, cycle-consistency assumes bijection, and CUT-style contrastive losses wrongly push apart subpatches with identical staining labels.
Approach
Decouple generator encoder features with an orthogonal MLP (Stiefel manifold constraint) and a learnable mask into staining-related and staining-unrelated components; apply a dual-stream PatchNCE where staining-related positives are all patches sharing the query's staining label, using pseudo-masks from UNI + ABMIL + Grad-CAM weak supervision; and align generated and real IHC feature distributions
Key results
- On BCI, ODA-GAN achieves KID 10.1 and FID 59.3 vs. ASP's 12.7 and 68.7; on MISTHER2 KID 8.6 vs. 23.9 for the best baseline ASP.
- On unpaired GPC3, ODA-GAN reaches FID 31.4 and KID 3.2, versus 39.4 and 4.7 for StegoGAN.
- The UNI-based pseudo-labeling with k=30 reaches patch-level AUC 0.933 and DICE 0.83 on GPC3, outperforming the PPIE baseline (AUC 0.782, DICE 0.71).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ODA-GAN instead of the strongest reported baseline | Virtual IHC staining · BCI | KID: 12.7 -> 10.1 | Helped |
| Use ODA-GAN instead of the strongest reported baseline | Virtual IHC staining · MISTHER2 | KID: 23.9 -> 8.6 | Helped |
| Use ODA-GAN instead of the strongest reported baseline | Virtual IHC staining · GPC3 | KID: 4.7 -> 3.2 | Helped |
| Add DPNCE to the CUT baseline (ablation) | ODA-GAN · GPC3 | KID: 25.3 -> 17.8 | Helped |
| Add MDA to the CUT baseline (ablation) | ODA-GAN · GPC3 | KID: 25.3 -> 12.0 | Helped |
| Add DPNCE and MDA to the CUT baseline (ablation) | ODA-GAN · GPC3 | KID: 25.3 -> 3.2 | Helped |
| Add DPNCE to the CUT baseline (ablation) | ODA-GAN · GPC3 | FID: 89.3 -> 60.2 | Helped |
| Add MDA to the CUT baseline (ablation) | ODA-GAN · GPC3 | FID: 89.3 -> 65.8 | Helped |
| Add DPNCE and MDA to the CUT baseline (ablation) | ODA-GAN · GPC3 | FID: 89.3 -> 31.4 | Helped |
| Set top-k selection to k=30 rather than k=10 or k=100 (ablation) | Our pseudo-labeling pipeline · GPC3 test set | AUC: 0.819 -> 0.933; 0.896 -> 0.933 | Helped |
| Use UNI-based pseudo-labeling instead of PPIE (ablation) | Patch-level classification and segmentation · GPC3 test set | AUC: 0.782 -> 0.933 | Helped |
| Use UNI-based pseudo-labeling instead of fine-tuning (ablation) | Patch-level classification and segmentation · GPC3 test set | AUC: 0.829 -> 0.933 | Helped |
Limitations
- The GPC3 dataset is private, limiting reproducibility of the unpaired results.
- On BCI, ODA-GAN's SSIM (0.4733) is second-best behind ASP (0.4917), and on MISTHER2 SSIM is also not best.
- Pseudo-label quality depends on the threshold and k in top-k selection; performance drops for k=10 or 100.
Experimental setup
Datasets: BCI MISTHER2 GPC3
Benchmarks: BCI MISTHER2 GPC3
Metrics: KID FID LPIPS SSIM DABKL AUC DICE IOU
Baselines: Pix2pix PyramidP2P TDKStain PSPStain ASP CycleGAN CUT DCD Confusion-GAN StegoGAN
Builds on: CUT (PatchNCE) CycleGAN UNI ABMIL Grad-CAM
Closest papers by meaning
- Virtual Immunohistochemistry Staining for Histological Images Assisted by Weakly-supervised Learning (CVPR 2024 · similarity 0.92)
- OT-StainNet: Optimal Transport Driven Semantic Matching for Weakly Paired H&E-to-IHC Stain Transfer (AAAI 2025 · similarity 0.89)
- Unpaired Multi-Domain Histopathology Virtual Staining Using Dual Path Prompted Inversion (AAAI 2025 · similarity 0.85)
- SheafStain: Sheaf-Theoretic Schr\"odinger Bridge for Spatially and Biologically Coherent Virtual Staining (NeurIPS 2026 · similarity 0.84)
- Multi-Resolution Pathology-Language Pre-training Model with Text-Guided Visual Representation (CVPR 2025 · similarity 0.84)
- BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational Pathology (CVPR 2025 · similarity 0.83)
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.