README.md

---
license: other
license_name: immunoinstruction-research-use
license_link: LICENSE
task_categories:
  - visual-question-answering
  - image-to-text
language:
  - en
tags:
  - pathology
  - histopathology
  - immunohistochemistry
  - computational-pathology
  - medical-imaging
  - vision-language
  - VQA
  - stain-transfer
pretty_name: ImmunoInstruction
size_categories:
  - 10K<n<100K
extra_gated_prompt: >-
  ## Terms of Access

  ImmunoInstruction is released for **non-commercial academic research
  purposes only**. By requesting access, you agree that:

  - You will use this dataset solely for research and educational purposes.
  - You will not attempt to re-identify any individual or institution
    associated with the underlying histopathology slides.
  - You will not redistribute the dataset, in whole or in part, without
    permission.
  - Any publication or derivative work using this dataset will cite the
    original paper (see Citation section).

  Please fill in the information below accurately. Access requests are
  reviewed manually.
extra_gated_fields:
  Full name: text
  Country: country
  Institution/Affiliation: text
  Email: text
  Intended use of this dataset: text
  I agree to use this dataset for non-commercial research purposes only: checkbox
extra_gated_button_content: "Submit access request"
extra_gated_heading: "Request access to ImmunoInstruction"
---

# ImmunoInstruction

**ImmunoInstruction** is a large-scale **IHC-positive-expression (IPE) instruction-following visual question answering (VQA) dataset** for computational pathology. It was constructed to train **VLEGA** (VL Expert-Guided Assessment model), a lightweight vision-language model that emulates pathologist reasoning to assess the quality of generated immunohistochemistry (IHC) images. VLEGA is the core evaluator behind **MEGFS** (Multimodal Expert-Guided Finer Selection), a key component of **DMCoStain**, an iterative data-model co-optimization framework for H&E-to-IHC stain transfer.

📄 Paper: *Towards Reliable Stain Transfer: An Iterative Data-Model Co-Optimization Framework Based on Multimodal Expert-Guided Assessment* (ACM Multimedia 2026)
💻 Code: https://github.com/SikangSHU/DMCoStain

## Dataset Summary

Images were collected from the training splits of two public H&E–IHC paired datasets, **MIST** and **HIT**, spanning two tissue types and six biomarkers:

| Subset | Tissue | Biomarker | Staining Location | # IHC Images |
|---|---|---|---|---|
| MIST-ER | Human breast cancer | ER | Nuclear | 4,093 |
| MIST-PR | Human breast cancer | PR | Nuclear | 4,134 |
| MIST-Ki67 | Human breast cancer | Ki67 | Nuclear | 4,334 |
| MIST-HER2 | Human breast cancer | HER2 | Membranous | 4,642 |
| HIT-PAX5 | Canine lymphoma | PAX5 | Nuclear | 6,043 |
| HIT-CD3 | Canine lymphoma | CD3 | Membranous | 6,228 |
| **Total** | | | | **29,474** |

Each IHC image (1024×1024 resolution) is paired with **4 question–answer pairs**, spanning four clinically grounded assessment categories defined by pathologists:

1. **Style Fidelity (C1)** — Whether the overall color, texture, and appearance of the image is consistent with real IHC images.
2. **Marker Location (C2)** — Which cellular compartment (membrane, cytoplasm, or nucleus) the positive signal is localized to.
3. **Marker Proportion & Spatial Position (C3)** — The approximate percentage of the image covered by the positive signal and its spatial distribution.
4. **Marker Intensity (C4)** — The strength of the positive signal (no signal / weak / moderate / strong).

> **Note on this release:** In the original paper, each image is additionally paired with a contrastive "Style Fidelity" negative question (C1-Q2), which uses out-of-domain noisy images (H&E, immunofluorescence, or heavily blurred IHC) to teach the model to reject non-IHC-style inputs. **These noisy negative images are not included in this release**, so each image here retains **4 QA pairs** (rather than 5) covering the four core categories above.

All initial answers were generated by GPT-4o following predefined instructions, then reviewed and refined by four pathology experts to ensure clinical accuracy.

## Dataset Structure

```
├── stage2_MIST_ER_png/       # IHC images (ER)
├── stage2_MIST_ER.json       # QA pairs (ER)
├── stage2_MIST_PR_png/
├── stage2_MIST_PR.json
├── stage2_MIST_Ki67_png/
├── stage2_MIST_Ki67.json
├── stage2_MIST_HER2_png/
├── stage2_MIST_HER2.json
├── stage2_HIT_PAX5_png/
├── stage2_HIT_PAX5.json
├── stage2_HIT_CD3_png/
└── stage2_HIT_CD3.json
```

Each `<subset>.json` file contains the QA annotations for the corresponding image folder, with each entry linking an image filename to its four question–answer pairs (Style Fidelity, Marker Location, Marker Proportion & Spatial Position, Marker Intensity). The `<subset>_csv/` folders provide the same annotations in tabular form.

## Data Splits

Following the original paper: for MIST-ER, MIST-PR, MIST-Ki67 (nuclear biomarkers) and HIT-CD3 (membranous), 80% of images are used for training and 20% for testing. MIST-HER2 (membranous) and HIT-PAX5 (nuclear) are held out entirely as **external test sets** to evaluate generalization to biomarkers unseen during training. Stratified sampling was used to preserve the natural (imbalanced) answer distribution within each biomarker.

## Intended Use

This dataset is intended for training and evaluating vision-language models on immunohistochemistry-positive-expression (IPE) assessment tasks — e.g., automated quality control of virtually generated (stain-transferred) IHC images, or general IHC image understanding. It underlies the VLEGA model used in the MEGFS/DMCoStain pipeline.

## Citation

If you use this dataset, please cite:

```bibtex
@article{xu2026towards,
  title={Towards Reliable Stain Transfer: An Iterative Data-Model Co-Optimization Framework Based on Multimodal Expert-Guided Assessment},
  author={Xu, Siyuan and Wang, Yan and Song, Haofei and Gao, Lili and Wang, Jiansheng and Zhang, Qing and Huang, Dan and Yun, Boxiang and Xiong, Hongkai and Li, Qingli},
  journal={arXiv preprint arXiv:2607.25393},
  year={2026}
}
```

## Source Data

Images are derived from the publicly released **MIST** [1] and **HIT** [2] datasets. Please also respect the original terms of use of MIST and HIT when using the underlying images.

[1] Li F, et al. Adaptive supervised patchnce loss for learning h&e-to-ihc stain translation with inconsistent groundtruth image pairs. International Conference on Medical Image Computing and Computer-Assisted Intervention. 2023: 632-641.

[2] Zhang W, et al. High-resolution medical image translation via patch alignment-based bidirectional contrastive learning. International Conference on Medical Image Computing and Computer-Assisted Intervention. 2024: 178-188.

## Contact

For questions about this dataset, please open a discussion on this Hugging Face repository or contact the first author listed in the paper.