Pathology foundation model for multi-stain virtual staining
| ISRCTN | ISRCTN13770753 |
|---|---|
| DOI | https://doi.org/10.1186/ISRCTN13770753 |
| Sponsor | First Affiliated Hospital of Jinan University |
| Funder | National Key Research and Development Program of China |
- Submission date
- 07/07/2026
- Registration date
- 08/07/2026
- Last edited
- 08/07/2026
- Recruitment status
- Recruiting
- Overall study status
- Ongoing
- Condition category
- Cancer
Plain English summary of protocol
Background and study aims
Pathologists usually stain tissue slides with chemicals (such as H&E or special stains for proteins) to see cell details and make diagnoses. However, staining takes time, costs money, and each stain often requires a separate piece of tissue. This study aims to develop a computer-based method that uses artificial intelligence to create digital images from unstained tissue slides – images that look just like the real stained ones. Our goal is to generate several different stain types from a single unstained slide while keeping the original tissue structure and cell details intact. We hope this virtual staining approach can work as well as, or even better than, conventional staining.
Who can participate?
Patients aged 18 years and over who have been diagnosed with one of the following cancers: nasopharyngeal, liver, breast, or colorectal cancer. Participants must have well-preserved tissue samples available for analysis.
What does the study involve?
If you agree to take part, we will use your existing tissue samples (already collected as part of your routine care) to test our AI model. No extra biopsies or procedures are required. The study involves taking unstained sections from your samples, scanning them with a digital scanner, and then running our deep learning algorithm to produce virtual stained images. We will compare these virtual images with the actual stained slides from your samples to see how closely they match. We will also ask clinicians to score the image quality and measure the diagnostic accuracy against the gold standard (real staining). The whole process is done on computer images – your tissue samples are handled according to standard laboratory safety protocols.
What are the possible benefits and risks of participating?
There is no direct clinical benefit to you personally from taking part. However, your participation will help us develop a new technique that could, in the future, save time, reduce costs, and preserve precious tissue for additional tests, potentially leading to faster and more accurate diagnoses for patients.
Since we only use leftover tissue from your routine care and do not perform any extra medical procedures, the risk to you is minimal. Your personal health information will be kept confidential and anonymised. There is no physical risk from the computer analysis itself.
Where is the study run from?
First Affiliated Hospital of Jinan University (China)
When is the study starting and how long is it expected to run for?
March 2026 to December 2026
Who is funding the study?
This research is funded by the National Key Research and Development Program of China (2023YFF1204600). The funder has no role in the study design, data collection, or interpretation of results.
Who is the main contact?
Dr Bin Zhang, xld_Jane_Eyre@126.com
Contact information
Principal investigator, Scientific, Public
No. 613 Huangpu West Road, Tianhe District
Guangzhou
510630
China
 ORCID ID |
0000-0002-6286-6227 |
|---|---|
| Phone | +86 (0)15217921427 |
| xld_Jane_Eyre@126.com |
Study information
| Primary study design | Observational |
|---|---|
| Observational study design | Cohort study |
| Scientific title | Pathology foundation model guided unified framework for multi-stain virtual staining |
| Study acronym | PFM-VS |
| Study objectives | This study maps an unstained formalin‑fixed paraffin‑embedded (FFPE) section to H&E and multiple immunohistochemical (IHC) or special stain appearances, while preserving the structural and pathological integrity of the tissue. |
| Ethics approval(s) |
Approved 25/03/2026, Ethics Committee of Jinan University (No. 601, Huangpu Avenue West, Tianhe District, Guangzhou, 510627, China; +86 (0)85220250; oykyc@jnu.edu.can), ref: JNUECKY-20260325-006 |
| Health condition(s) or problem(s) studied | Nasopharyngeal carcinoma, hepatocellular carcinoma, breast cancer, colorectal cancer, or other malignant tumors |
| Methodology | Existing virtual staining methods are typically formulated as independent image‑to‑image translation tasks, in which a separate model is trained for each target stain. Although such approaches can perform specific conversions - for instance, generating H&E images from unstained sections, or producing a particular immunohistochemical stain from unstained ones - they inherently treat different staining tasks as isolated problems. This modelling strategy neglects the shared foundational pathological information that underlies both histological and immunohistochemical staining. In clinical histopathological diagnosis, H&E and multiple IHC stains represent different visualisations of the same tissue state, collectively reflecting pathological features such as tissue architecture, cellular composition, nuclear morphology, and the local microenvironment. In this work, we formulate multi‑stain virtual staining as a unified conditional generation problem and propose a pathology‑foundation‑model‑based unified framework that generates multiple target stains from a single unstained tissue image within a single conditional generative model. The core design principle of our framework is to decouple stain‑agnostic shared pathological information from stain‑specific visual appearances. The model first extracts a shared pathological representation from the unstained image, which captures tissue architecture, cell morphology, and spatial microenvironmental cues that are common to different stains. Then, the target stain condition modulates this shared representation, allowing the generative model to selectively produce the visual features and biological patterns characteristic of the desired stain. In this way, the model can learn a common pathological basis across diverse staining tasks while preserving the distinct appearance and expression profiles of each stain. |
| Intervention type | Other |
| Primary outcome measure(s) |
|
| Key secondary outcome measure(s) | |
| Completion date | 31/12/2026 |
Eligibility
| Participant type(s) | |
|---|---|
| Age group | Mixed |
| Lower age limit | 18 Years |
| Upper age limit | 100 Years |
| Sex | All |
| Target sample size at registration | 400 |
| Total final enrolment | 400 |
| Key inclusion criteria | 1. Age ≥ 18 years 2. Patients diagnosed with nasopharyngeal carcinoma, hepatocellular carcinoma, breast cancer, colorectal cancer, or other malignant tumors 3. Well-preserved formalin‑fixed paraffin‑embedded (FFPE) tissue samples |
| Key exclusion criteria | 1. Age <18 years 2. Not diagnosed with malignant tumors, or diagnosis is unclear 3. Insufficient FFPE tissue samples, or samples with poor quality that do not meet testing requirements 4. H&E and IHC staining are unsatisfactory (e.g., poor section quality, artifacts, or insufficient cellularity) |
| Date of first enrolment | 26/03/2026 |
| Date of final enrolment | 30/09/2026 |
Locations
Countries of recruitment
- China
Study participating centres
Results and Publications
| Individual participant data (IPD) Intention to share | No |
|---|
Editorial Notes
07/07/2026: Study's existence confirmed by the Ethics Committee of Jinan University.
